diff --git a/.idea/CASD.iml b/.idea/CASD.iml
new file mode 100644
index 0000000..7c9d48f
--- /dev/null
+++ b/.idea/CASD.iml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="PYTHON_MODULE" version="4">
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$" />
+    <orderEntry type="inheritedJdk" />
+    <orderEntry type="sourceFolder" forTests="false" />
+  </component>
+  <component name="TestRunnerService">
+    <option name="projectConfiguration" value="pytest" />
+    <option name="PROJECT_TEST_RUNNER" value="pytest" />
+  </component>
+</module>
\ No newline at end of file
diff --git a/.idea/encodings.xml b/.idea/encodings.xml
new file mode 100644
index 0000000..15a15b2
--- /dev/null
+++ b/.idea/encodings.xml
@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="Encoding" addBOMForNewFiles="with NO BOM" />
+</project>
\ No newline at end of file
diff --git a/.idea/misc.xml b/.idea/misc.xml
new file mode 100644
index 0000000..cf7d47d
--- /dev/null
+++ b/.idea/misc.xml
@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="JavaScriptSettings">
+    <option name="languageLevel" value="ES6" />
+  </component>
+  <component name="ProjectRootManager" version="2" project-jdk-name="Python 3.7 (view_attention)" project-jdk-type="Python SDK" />
+</project>
\ No newline at end of file
diff --git a/.idea/modules.xml b/.idea/modules.xml
new file mode 100644
index 0000000..99d4ae6
--- /dev/null
+++ b/.idea/modules.xml
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/CASD.iml" filepath="$PROJECT_DIR$/.idea/CASD.iml" />
+    </modules>
+  </component>
+</project>
\ No newline at end of file
diff --git a/.idea/vcs.xml b/.idea/vcs.xml
new file mode 100644
index 0000000..94a25f7
--- /dev/null
+++ b/.idea/vcs.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="VcsDirectoryMappings">
+    <mapping directory="$PROJECT_DIR$" vcs="Git" />
+  </component>
+</project>
\ No newline at end of file
diff --git a/.idea/workspace.xml b/.idea/workspace.xml
new file mode 100644
index 0000000..01df77f
--- /dev/null
+++ b/.idea/workspace.xml
@@ -0,0 +1,259 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ChangeListManager">
+    <list default="true" id="e3fd67ad-91e8-4254-87ab-c83aca39f8d3" name="Default Changelist" comment="" />
+    <option name="EXCLUDED_CONVERTED_TO_IGNORED" value="true" />
+    <option name="SHOW_DIALOG" value="false" />
+    <option name="HIGHLIGHT_CONFLICTS" value="true" />
+    <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
+    <option name="LAST_RESOLUTION" value="IGNORE" />
+  </component>
+  <component name="FileEditorManager">
+    <leaf SIDE_TABS_SIZE_LIMIT_KEY="300">
+      <file pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/tools/trainval_net.py">
+          <provider selected="true" editor-type-id="text-editor">
+            <state relative-caret-position="1422">
+              <caret line="94" selection-start-line="94" selection-end-line="94" />
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file pinned="false" current-in-tab="true">
+        <entry file="file://$PROJECT_DIR$/lib/nets/network.py">
+          <provider selected="true" editor-type-id="text-editor">
+            <state relative-caret-position="3865">
+              <caret line="258" column="50" selection-start-line="258" selection-start-column="50" selection-end-line="258" selection-end-column="50" />
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/tools/test_net.py">
+          <provider selected="true" editor-type-id="text-editor">
+            <state relative-caret-position="1854">
+              <caret line="120" column="28" selection-start-line="120" selection-start-column="28" selection-end-line="120" selection-end-column="28" />
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/lib/model/config.py">
+          <provider selected="true" editor-type-id="text-editor">
+            <state relative-caret-position="5400">
+              <caret line="308" column="32" selection-start-line="308" selection-start-column="32" selection-end-line="308" selection-end-column="32" />
+              <folding>
+                <element signature="e#0#38#0" expanded="true" />
+              </folding>
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/lib/model/train_val.py">
+          <provider selected="true" editor-type-id="text-editor">
+            <state relative-caret-position="4248">
+              <caret line="245" column="35" selection-start-line="245" selection-start-column="35" selection-end-line="245" selection-end-column="35" />
+            </state>
+          </provider>
+        </entry>
+      </file>
+    </leaf>
+  </component>
+  <component name="FindInProjectRecents">
+    <findStrings>
+      <find>self._losses['cls_det_loss']</find>
+    </findStrings>
+  </component>
+  <component name="Git.Settings">
+    <option name="RECENT_GIT_ROOT_PATH" value="$PROJECT_DIR$" />
+  </component>
+  <component name="IdeDocumentHistory">
+    <option name="CHANGED_PATHS">
+      <list>
+        <option value="$PROJECT_DIR$/lib/model/config.py" />
+        <option value="$PROJECT_DIR$/lib/nets/network.py" />
+        <option value="$PROJECT_DIR$/tools/demo.py" />
+        <option value="$PROJECT_DIR$/tools/test_net.py" />
+        <option value="$PROJECT_DIR$/tools/trainval_net.py" />
+      </list>
+    </option>
+  </component>
+  <component name="ProjectFrameBounds" extendedState="6">
+    <option name="x" value="65" />
+    <option name="y" value="-4" />
+    <option name="width" value="1855" />
+    <option name="height" value="1084" />
+  </component>
+  <component name="ProjectView">
+    <navigator proportions="" version="1">
+      <foldersAlwaysOnTop value="true" />
+    </navigator>
+    <panes>
+      <pane id="Scope" />
+      <pane id="ProjectPane">
+        <subPane>
+          <expand>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="lib" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="lib" type="462c0819:PsiDirectoryNode" />
+              <item name="datasets" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="lib" type="462c0819:PsiDirectoryNode" />
+              <item name="model" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="lib" type="462c0819:PsiDirectoryNode" />
+              <item name="nets" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="output" type="462c0819:PsiDirectoryNode" />
+            </path>
+            <path>
+              <item name="CASD" type="b2602c69:ProjectViewProjectNode" />
+              <item name="CASD" type="462c0819:PsiDirectoryNode" />
+              <item name="tools" type="462c0819:PsiDirectoryNode" />
+            </path>
+          </expand>
+          <select />
+        </subPane>
+      </pane>
+    </panes>
+  </component>
+  <component name="PropertiesComponent">
+    <property name="WebServerToolWindowFactoryState" value="true" />
+    <property name="last_opened_file_path" value="$PROJECT_DIR$" />
+    <property name="nodejs_interpreter_path.stuck_in_default_project" value="undefined stuck path" />
+    <property name="nodejs_npm_path_reset_for_default_project" value="true" />
+  </component>
+  <component name="RunDashboard">
+    <option name="ruleStates">
+      <list>
+        <RuleState>
+          <option name="name" value="ConfigurationTypeDashboardGroupingRule" />
+        </RuleState>
+        <RuleState>
+          <option name="name" value="StatusDashboardGroupingRule" />
+        </RuleState>
+      </list>
+    </option>
+  </component>
+  <component name="SvnConfiguration">
+    <configuration />
+  </component>
+  <component name="TaskManager">
+    <task active="true" id="Default" summary="Default task">
+      <changelist id="e3fd67ad-91e8-4254-87ab-c83aca39f8d3" name="Default Changelist" comment="" />
+      <created>1607204328623</created>
+      <option name="number" value="Default" />
+      <option name="presentableId" value="Default" />
+      <updated>1607204328623</updated>
+      <workItem from="1607204329974" duration="783000" />
+      <workItem from="1607205646014" duration="11000" />
+    </task>
+    <servers />
+  </component>
+  <component name="TimeTrackingManager">
+    <option name="totallyTimeSpent" value="794000" />
+  </component>
+  <component name="ToolWindowManager">
+    <frame x="65" y="-4" width="1855" height="1084" extended-state="6" />
+    <layout>
+      <window_info active="true" content_ui="combo" id="Project" order="0" visible="true" weight="0.20619127" />
+      <window_info id="Structure" order="1" side_tool="true" weight="0.25" />
+      <window_info id="Favorites" order="2" side_tool="true" />
+      <window_info anchor="bottom" id="Message" order="0" />
+      <window_info anchor="bottom" id="Find" order="1" />
+      <window_info anchor="bottom" id="Run" order="2" />
+      <window_info anchor="bottom" id="Debug" order="3" weight="0.4" />
+      <window_info anchor="bottom" id="Cvs" order="4" weight="0.25" />
+      <window_info anchor="bottom" id="Inspection" order="5" weight="0.4" />
+      <window_info anchor="bottom" id="TODO" order="6" />
+      <window_info anchor="bottom" id="Docker" order="7" show_stripe_button="false" />
+      <window_info anchor="bottom" id="Version Control" order="8" />
+      <window_info anchor="bottom" id="Database Changes" order="9" />
+      <window_info anchor="bottom" id="Event Log" order="10" side_tool="true" />
+      <window_info anchor="bottom" id="Terminal" order="11" />
+      <window_info anchor="bottom" id="Python Console" order="12" />
+      <window_info anchor="right" id="Commander" internal_type="SLIDING" order="0" type="SLIDING" weight="0.4" />
+      <window_info anchor="right" id="Ant Build" order="1" weight="0.25" />
+      <window_info anchor="right" content_ui="combo" id="Hierarchy" order="2" weight="0.25" />
+      <window_info anchor="right" id="Remote Host" order="3" />
+      <window_info anchor="right" id="SciView" order="4" />
+      <window_info anchor="right" id="Database" order="5" />
+    </layout>
+  </component>
+  <component name="TypeScriptGeneratedFilesManager">
+    <option name="version" value="1" />
+  </component>
+  <component name="editorHistoryManager">
+    <entry file="file://$PROJECT_DIR$/tools/_init_paths.py">
+      <provider selected="true" editor-type-id="text-editor" />
+    </entry>
+    <entry file="file://$PROJECT_DIR$/lib/nets/vgg16.py">
+      <provider selected="true" editor-type-id="text-editor" />
+    </entry>
+    <entry file="file://$PROJECT_DIR$/tools/demo.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="-379">
+          <caret line="40" column="48" lean-forward="true" selection-start-line="40" selection-start-column="48" selection-end-line="40" selection-end-column="48" />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/tools/trainval_net.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="1422">
+          <caret line="94" selection-start-line="94" selection-end-line="94" />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/tools/test_net.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="1854">
+          <caret line="120" column="28" selection-start-line="120" selection-start-column="28" selection-end-line="120" selection-end-column="28" />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/lib/model/config.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="5400">
+          <caret line="308" column="32" selection-start-line="308" selection-start-column="32" selection-end-line="308" selection-end-column="32" />
+          <folding>
+            <element signature="e#0#38#0" expanded="true" />
+          </folding>
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/lib/model/train_val.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="4248">
+          <caret line="245" column="35" selection-start-line="245" selection-start-column="35" selection-end-line="245" selection-end-column="35" />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/lib/nets/network.py">
+      <provider selected="true" editor-type-id="text-editor">
+        <state relative-caret-position="3865">
+          <caret line="258" column="50" selection-start-line="258" selection-start-column="50" selection-end-line="258" selection-end-column="50" />
+        </state>
+      </provider>
+    </entry>
+  </component>
+</project>
\ No newline at end of file
diff --git a/experiments/cfgs/change_log.txt b/experiments/cfgs/change_log.txt
new file mode 100644
index 0000000..23d68d5
--- /dev/null
+++ b/experiments/cfgs/change_log.txt
@@ -0,0 +1,48 @@
+2018/11/22/9:37
+In the vgg16.yml files, The orginal TEST.NMS value is 0.3 but changed to 0.4 for 
+implementation for WSDNN 
+
+2018/11/23/14:19
+In vgg16.yml files, the original POOLING_MODE == crop, however changed to roi_align
+for implementation for WSDNN
+
+2018/11/24/8:37
+In vgg16.yml file, the original TRAIN.WEIGHT_DECAY == 0.0001, and changed to 0.005
+Also ss_boxes whose widths or heights are less than 20 are removed
+
+2018/11/24/9:01
+In vgg16.yml files, the TRAIN.SCALES is {600,}, however changed to {480, 576, 688, 864, 1200}
+TEST.SCLAES is {688, }
+
+2018/11/24/9:23
+In vgg16.yml file, the TRAIN.MAX_SIZE is 1000, however changed to 1200
+
+2018/11/24/20:33
+In lib/datasets/pascal_voc.py, changed classes from 21 classes to 20 classes
+original:
+self._classes = ('__background__',  # always index 0
+                     'aeroplane', 'bicycle', 'bird', 'boat',
+                     'bottle', 'bus', 'car', 'cat', 'chair',
+                     'cow', 'diningtable', 'dog', 'horse',
+                     'motorbike', 'person', 'pottedplant',
+                     'sheep', 'sofa', 'train', 'tvmonitor')
+new:
+self._classes = (
+                     'aeroplane', 'bicycle', 'bird', 'boat',
+                     'bottle', 'bus', 'car', 'cat', 'chair',
+                     'cow', 'diningtable', 'dog', 'horse',
+                     'motorbike', 'person', 'pottedplant',
+                     'sheep', 'sofa', 'train', 'tvmonitor')
+In network.py, added a score_det_net
+
+In vgg16.yml file, changed TRAIN.STEPSIZE from 30000 to 60000
+
+
+2018/11/25/9:59
+In vgg16.yml, the learning rate is 0.0001, and changed to 0.0005.
+
+2018/11/26/11:14
+In vgg16.yml, the TEST.SCLAES changed from [688] to [480, 576, 688, 864, 1200]
+for multiple scale test
+
+
diff --git a/experiments/cfgs/mobile.yml b/experiments/cfgs/mobile.yml
new file mode 100644
index 0000000..8adbf94
--- /dev/null
+++ b/experiments/cfgs/mobile.yml
@@ -0,0 +1,16 @@
+EXP_DIR: mobile
+TRAIN:
+  HAS_RPN: True
+  IMS_PER_BATCH: 1
+  BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
+  RPN_POSITIVE_OVERLAP: 0.7
+  RPN_BATCHSIZE: 256
+  PROPOSAL_METHOD: gt
+  BG_THRESH_LO: 0.0
+  DISPLAY: 20
+  BATCH_SIZE: 256
+  DOUBLE_BIAS: False
+  SNAPSHOT_PREFIX: mobile_faster_rcnn
+TEST:
+  HAS_RPN: True
+POOLING_MODE: crop
diff --git a/experiments/cfgs/res101-lg.yml b/experiments/cfgs/res101-lg.yml
new file mode 100644
index 0000000..ce4aefb
--- /dev/null
+++ b/experiments/cfgs/res101-lg.yml
@@ -0,0 +1,22 @@
+EXP_DIR: res101-lg
+TRAIN:
+  HAS_RPN: True
+  IMS_PER_BATCH: 1
+  BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
+  RPN_POSITIVE_OVERLAP: 0.7
+  RPN_BATCHSIZE: 256
+  PROPOSAL_METHOD: gt
+  BG_THRESH_LO: 0.0
+  DISPLAY: 20
+  BATCH_SIZE: 256
+  DOUBLE_BIAS: False
+  SNAPSHOT_PREFIX: res101_faster_rcnn
+  SCALES: [800]
+  MAX_SIZE: 1333
+TEST:
+  HAS_RPN: True
+  SCALES: [800]
+  MAX_SIZE: 1333
+  RPN_POST_NMS_TOP_N: 1000
+POOLING_MODE: crop
+ANCHOR_SCALES: [2,4,8,16,32]
diff --git a/experiments/cfgs/res101.yml b/experiments/cfgs/res101.yml
new file mode 100644
index 0000000..367cad3
--- /dev/null
+++ b/experiments/cfgs/res101.yml
@@ -0,0 +1,16 @@
+EXP_DIR: res101
+TRAIN:
+  HAS_RPN: True
+  IMS_PER_BATCH: 1
+  BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
+  RPN_POSITIVE_OVERLAP: 0.7
+  RPN_BATCHSIZE: 256
+  PROPOSAL_METHOD: gt
+  BG_THRESH_LO: 0.0
+  DISPLAY: 20
+  BATCH_SIZE: 256
+  DOUBLE_BIAS: False
+  SNAPSHOT_PREFIX: res101_faster_rcnn
+TEST:
+  HAS_RPN: True
+POOLING_MODE: crop
diff --git a/experiments/cfgs/res50.yml b/experiments/cfgs/res50.yml
new file mode 100644
index 0000000..7b38b2a
--- /dev/null
+++ b/experiments/cfgs/res50.yml
@@ -0,0 +1,16 @@
+EXP_DIR: res50
+TRAIN:
+  HAS_RPN: True
+  IMS_PER_BATCH: 1
+  BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
+  RPN_POSITIVE_OVERLAP: 0.7
+  RPN_BATCHSIZE: 256
+  PROPOSAL_METHOD: gt
+  BG_THRESH_LO: 0.0
+  DISPLAY: 20
+  BATCH_SIZE: 256
+  DOUBLE_BIAS: False
+  SNAPSHOT_PREFIX: res50_faster_rcnn
+TEST:
+  HAS_RPN: True
+POOLING_MODE: crop
diff --git a/experiments/cfgs/vgg16.yml b/experiments/cfgs/vgg16.yml
new file mode 100644
index 0000000..d13ede6
--- /dev/null
+++ b/experiments/cfgs/vgg16.yml
@@ -0,0 +1,25 @@
+EXP_DIR: vgg16_MELM
+TRAIN:
+  HAS_RPN: False
+  IMS_PER_BATCH: 1
+  BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
+  RPN_POSITIVE_OVERLAP: 0.7
+  RPN_BATCHSIZE: 256
+  PROPOSAL_METHOD: selective_search
+  BG_THRESH_LO: 0.0
+  DISPLAY: 20
+  BATCH_SIZE: 256
+  SNAPSHOT_PREFIX: vgg16_MELM
+  LEARNING_RATE: 0.001  # 0.001
+  WEIGHT_DECAY: 0.0005
+  SCALES: [480, 576, 688, 864, 1200]
+  MAX_SIZE: 2000
+  STEPSIZE: [50000,]
+  MIL_RECURRENT_STEP: 20000
+  MIL_RECURRECT_WEIGHT: 0.09
+TEST:
+  HAS_RPN: False
+  PROPOSAL_METHOD: selective_search
+  NMS: 0.3
+  SCALES: [480, 576, 688, 864, 1200]
+POOLING_MODE: roi_align
diff --git a/experiments/scripts/convert_vgg16.sh b/experiments/scripts/convert_vgg16.sh
new file mode 100644
index 0000000..7fae1b4
--- /dev/null
+++ b/experiments/scripts/convert_vgg16.sh
@@ -0,0 +1,67 @@
+#!/bin/bash
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=vgg16
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:2:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=70000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+set +x
+NET_FINAL=${NET}_faster_rcnn_iter_${ITERS}
+set -x
+
+if [ ! -f ${NET_FINAL}.index ]; then
+    if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+        CUDA_VISIBLE_DEVICES=${GPU_ID} time python ./tools/convert_from_depre.py \
+            --snapshot ${NET_FINAL} \
+            --imdb ${TRAIN_IMDB} \
+            --iters ${ITERS} \
+            --cfg experiments/cfgs/${NET}.yml \
+            --tag ${EXTRA_ARGS_SLUG} \
+            --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} ${EXTRA_ARGS}
+    else
+        CUDA_VISIBLE_DEVICES=${GPU_ID} time python ./tools/convert_from_depre.py \
+            --snapshot ${NET_FINAL} \
+            --imdb ${TRAIN_IMDB} \
+            --iters ${ITERS} \
+            --cfg experiments/cfgs/${NET}.yml \
+            --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} ${EXTRA_ARGS}
+    fi
+fi
+
diff --git a/experiments/scripts/test_faster_rcnn.sh b/experiments/scripts/test_faster_rcnn.sh
new file mode 100644
index 0000000..d613656
--- /dev/null
+++ b/experiments/scripts/test_faster_rcnn.sh
@@ -0,0 +1,76 @@
+#!/bin/bash
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=100000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/test_${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}_MELM/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_MELM_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}_MELM/${TRAIN_IMDB}/default/${NET}_MELM_iter_${ITERS}.pth
+
+fi
+set -x
+
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/test_net.py \
+    --imdb ${TEST_IMDB} \
+    --model ${NET_FINAL} \
+    --cfg experiments/cfgs/${NET}.yml \
+    --tag ${EXTRA_ARGS_SLUG} \
+    --net ${NET} \
+    --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+          ${EXTRA_ARGS}
+else
+  CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/test_net.py \
+    --imdb ${TEST_IMDB} \
+    --model ${NET_FINAL} \
+    --cfg experiments/cfgs/${NET}.yml \
+    --net ${NET} \
+    --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+          ${EXTRA_ARGS}
+fi
+
diff --git a/experiments/scripts/test_faster_rcnn_notime.sh b/experiments/scripts/test_faster_rcnn_notime.sh
new file mode 100644
index 0000000..53f00e3
--- /dev/null
+++ b/experiments/scripts/test_faster_rcnn_notime.sh
@@ -0,0 +1,75 @@
+#!/bin/bash
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=70000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/test_${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_faster_rcnn_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/default/${NET}_faster_rcnn_iter_${ITERS}.pth
+fi
+set -x
+
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/test_net.py \
+    --imdb ${TEST_IMDB} \
+    --model ${NET_FINAL} \
+    --cfg experiments/cfgs/${NET}.yml \
+    --tag ${EXTRA_ARGS_SLUG} \
+    --net ${NET} \
+    --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+          ${EXTRA_ARGS}
+else
+  CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/test_net.py \
+    --imdb ${TEST_IMDB} \
+    --model ${NET_FINAL} \
+    --cfg experiments/cfgs/${NET}.yml \
+    --net ${NET} \
+    --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+          ${EXTRA_ARGS}
+fi
+
diff --git a/experiments/scripts/train_faster_rcnn.sh b/experiments/scripts/train_faster_rcnn.sh
new file mode 100644
index 0000000..ca1a081
--- /dev/null
+++ b/experiments/scripts/train_faster_rcnn.sh
@@ -0,0 +1,88 @@
+#!/bin/bash
+
+alias time='/usr/bin/time'
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[50000]"
+    ITERS=100000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[80000]"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    STEPSIZE="[350000]"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}_${NET}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_MELM_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/default/${NET}_MELM_iter_${ITERS}.pth
+fi
+set -x
+
+if [ ! -f ${NET_FINAL}.index ]; then
+  if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --tag ${EXTRA_ARGS_SLUG} \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  else
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  fi
+fi
+
+echo $@
+./experiments/scripts/test_faster_rcnn.sh $@
diff --git a/experiments/scripts/train_faster_rcnn_notime.sh b/experiments/scripts/train_faster_rcnn_notime.sh
new file mode 100644
index 0000000..eddd772
--- /dev/null
+++ b/experiments/scripts/train_faster_rcnn_notime.sh
@@ -0,0 +1,85 @@
+#!/bin/bash
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[50000]"
+    ITERS=70000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[80000]"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    STEPSIZE="[900000]"
+    ITERS=1190000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}_${NET}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_faster_rcnn_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/default/${NET}_faster_rcnn_iter_${ITERS}.pth
+fi
+set -x
+
+if [ ! -f ${NET_FINAL}.index ]; then
+  if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --tag ${EXTRA_ARGS_SLUG} \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  else
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  fi
+fi
+
+./experiments/scripts/test_faster_rcnn_notime.sh $@
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..64e309d
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..f3e785a
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..39168d0
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..a5d45b7
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..f4bcc00
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..7e4de05
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..25aed8a
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/lib.linux-x86_64-3.7/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/build/lib.linux-x86_64-3.7/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..69ee1a3
Binary files /dev/null and b/lib/build/lib.linux-x86_64-3.7/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cpu.o b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cpu.o
new file mode 100644
index 0000000..a657583
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cpu.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cuda.o b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cuda.o
new file mode 100644
index 0000000..85fd57c
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_cuda.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_kernel.o b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_kernel.o
new file mode 100644
index 0000000..3e3b755
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/nms_kernel.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/nms/src/soft_nms_cpu.o b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/soft_nms_cpu.o
new file mode 100644
index 0000000..655e012
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/nms/src/soft_nms_cpu.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_cuda.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_cuda.o
new file mode 100644
index 0000000..9a0771f
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_cuda.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_kernel.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_kernel.o
new file mode 100644
index 0000000..b75ed35
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_align/src/roi_align_kernel.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cpu.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cpu.o
new file mode 100644
index 0000000..36c508b
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cpu.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cuda.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cuda.o
new file mode 100644
index 0000000..890af9b
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_cuda.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_kernel.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_kernel.o
new file mode 100644
index 0000000..3e568d7
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_crop/src/roi_crop_kernel.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_cuda.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_cuda.o
new file mode 100644
index 0000000..3df3ee6
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_cuda.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_kernel.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_kernel.o
new file mode 100644
index 0000000..c9c9a61
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_pool/src/roi_pool_kernel.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_cuda.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_cuda.o
new file mode 100644
index 0000000..53d8a5d
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_cuda.o differ
diff --git a/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_kernel.o b/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_kernel.o
new file mode 100644
index 0000000..810d532
Binary files /dev/null and b/lib/build/temp.linux-x86_64-3.7/ops/roi_ring_pool/src/roi_ring_pool_kernel.o differ
diff --git a/lib/datasets/VOCdevkit-matlab-wrapper/get_voc_opts.m b/lib/datasets/VOCdevkit-matlab-wrapper/get_voc_opts.m
new file mode 100644
index 0000000..629597a
--- /dev/null
+++ b/lib/datasets/VOCdevkit-matlab-wrapper/get_voc_opts.m
@@ -0,0 +1,14 @@
+function VOCopts = get_voc_opts(path)
+
+tmp = pwd;
+cd(path);
+try
+  addpath('VOCcode');
+  VOCinit;
+catch
+  rmpath('VOCcode');
+  cd(tmp);
+  error(sprintf('VOCcode directory not found under %s', path));
+end
+rmpath('VOCcode');
+cd(tmp);
diff --git a/lib/datasets/VOCdevkit-matlab-wrapper/voc_eval.m b/lib/datasets/VOCdevkit-matlab-wrapper/voc_eval.m
new file mode 100644
index 0000000..1911a0e
--- /dev/null
+++ b/lib/datasets/VOCdevkit-matlab-wrapper/voc_eval.m
@@ -0,0 +1,56 @@
+function res = voc_eval(path, comp_id, test_set, output_dir)
+
+VOCopts = get_voc_opts(path);
+VOCopts.testset = test_set;
+
+for i = 1:length(VOCopts.classes)
+  cls = VOCopts.classes{i};
+  res(i) = voc_eval_cls(cls, VOCopts, comp_id, output_dir);
+end
+
+fprintf('\n~~~~~~~~~~~~~~~~~~~~\n');
+fprintf('Results:\n');
+aps = [res(:).ap]';
+fprintf('%.1f\n', aps * 100);
+fprintf('%.1f\n', mean(aps) * 100);
+fprintf('~~~~~~~~~~~~~~~~~~~~\n');
+
+function res = voc_eval_cls(cls, VOCopts, comp_id, output_dir)
+
+test_set = VOCopts.testset;
+year = VOCopts.dataset(4:end);
+
+addpath(fullfile(VOCopts.datadir, 'VOCcode'));
+
+res_fn = sprintf(VOCopts.detrespath, comp_id, cls);
+
+recall = [];
+prec = [];
+ap = 0;
+ap_auc = 0;
+
+do_eval = (str2num(year) <= 2007) | ~strcmp(test_set, 'test');
+if do_eval
+  % Bug in VOCevaldet requires that tic has been called first
+  tic;
+  [recall, prec, ap] = VOCevaldet(VOCopts, comp_id, cls, true);
+  ap_auc = xVOCap(recall, prec);
+
+  % force plot limits
+  ylim([0 1]);
+  xlim([0 1]);
+
+  print(gcf, '-djpeg', '-r0', ...
+        [output_dir '/' cls '_pr.jpg']);
+end
+fprintf('!!! %s : %.4f %.4f\n', cls, ap, ap_auc);
+
+res.recall = recall;
+res.prec = prec;
+res.ap = ap;
+res.ap_auc = ap_auc;
+
+save([output_dir '/' cls '_pr.mat'], ...
+     'res', 'recall', 'prec', 'ap', 'ap_auc');
+
+rmpath(fullfile(VOCopts.datadir, 'VOCcode'));
diff --git a/lib/datasets/VOCdevkit-matlab-wrapper/xVOCap.m b/lib/datasets/VOCdevkit-matlab-wrapper/xVOCap.m
new file mode 100644
index 0000000..de6c628
--- /dev/null
+++ b/lib/datasets/VOCdevkit-matlab-wrapper/xVOCap.m
@@ -0,0 +1,10 @@
+function ap = xVOCap(rec,prec)
+% From the PASCAL VOC 2011 devkit
+
+mrec=[0 ; rec ; 1];
+mpre=[0 ; prec ; 0];
+for i=numel(mpre)-1:-1:1
+    mpre(i)=max(mpre(i),mpre(i+1));
+end
+i=find(mrec(2:end)~=mrec(1:end-1))+1;
+ap=sum((mrec(i)-mrec(i-1)).*mpre(i));
diff --git a/lib/datasets/__init__.py b/lib/datasets/__init__.py
new file mode 100644
index 0000000..7ba6a65
--- /dev/null
+++ b/lib/datasets/__init__.py
@@ -0,0 +1,6 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
diff --git a/lib/datasets/__pycache__/__init__.cpython-36.pyc b/lib/datasets/__pycache__/__init__.cpython-36.pyc
new file mode 100644
index 0000000..1c88744
Binary files /dev/null and b/lib/datasets/__pycache__/__init__.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/__init__.cpython-37.pyc b/lib/datasets/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..46d1ce0
Binary files /dev/null and b/lib/datasets/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/coco.cpython-36.pyc b/lib/datasets/__pycache__/coco.cpython-36.pyc
new file mode 100644
index 0000000..e35f6bb
Binary files /dev/null and b/lib/datasets/__pycache__/coco.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/coco.cpython-37.pyc b/lib/datasets/__pycache__/coco.cpython-37.pyc
new file mode 100644
index 0000000..469143c
Binary files /dev/null and b/lib/datasets/__pycache__/coco.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/ds_utils.cpython-36.pyc b/lib/datasets/__pycache__/ds_utils.cpython-36.pyc
new file mode 100644
index 0000000..223a180
Binary files /dev/null and b/lib/datasets/__pycache__/ds_utils.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/ds_utils.cpython-37.pyc b/lib/datasets/__pycache__/ds_utils.cpython-37.pyc
new file mode 100644
index 0000000..dd7faeb
Binary files /dev/null and b/lib/datasets/__pycache__/ds_utils.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/factory.cpython-36.pyc b/lib/datasets/__pycache__/factory.cpython-36.pyc
new file mode 100644
index 0000000..9326172
Binary files /dev/null and b/lib/datasets/__pycache__/factory.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/factory.cpython-37.pyc b/lib/datasets/__pycache__/factory.cpython-37.pyc
new file mode 100644
index 0000000..748a431
Binary files /dev/null and b/lib/datasets/__pycache__/factory.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/imdb.cpython-36.pyc b/lib/datasets/__pycache__/imdb.cpython-36.pyc
new file mode 100644
index 0000000..545cfe1
Binary files /dev/null and b/lib/datasets/__pycache__/imdb.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/imdb.cpython-37.pyc b/lib/datasets/__pycache__/imdb.cpython-37.pyc
new file mode 100644
index 0000000..2069294
Binary files /dev/null and b/lib/datasets/__pycache__/imdb.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/pascal_voc.cpython-36.pyc b/lib/datasets/__pycache__/pascal_voc.cpython-36.pyc
new file mode 100644
index 0000000..559fb14
Binary files /dev/null and b/lib/datasets/__pycache__/pascal_voc.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/pascal_voc.cpython-37.pyc b/lib/datasets/__pycache__/pascal_voc.cpython-37.pyc
new file mode 100644
index 0000000..9ba1390
Binary files /dev/null and b/lib/datasets/__pycache__/pascal_voc.cpython-37.pyc differ
diff --git a/lib/datasets/__pycache__/voc_eval.cpython-36.pyc b/lib/datasets/__pycache__/voc_eval.cpython-36.pyc
new file mode 100644
index 0000000..c8147d2
Binary files /dev/null and b/lib/datasets/__pycache__/voc_eval.cpython-36.pyc differ
diff --git a/lib/datasets/__pycache__/voc_eval.cpython-37.pyc b/lib/datasets/__pycache__/voc_eval.cpython-37.pyc
new file mode 100644
index 0000000..f899d4a
Binary files /dev/null and b/lib/datasets/__pycache__/voc_eval.cpython-37.pyc differ
diff --git a/lib/datasets/coco.py b/lib/datasets/coco.py
new file mode 100644
index 0000000..fba89f0
--- /dev/null
+++ b/lib/datasets/coco.py
@@ -0,0 +1,316 @@
+# --------------------------------------------------------
+# Fast/er R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick and Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from datasets.imdb import imdb
+import datasets.ds_utils as ds_utils
+from model.config import cfg
+import os.path as osp
+import sys
+import os
+import numpy as np
+import scipy.sparse
+import scipy.io as sio
+import pickle
+import json
+import uuid
+# COCO API
+from pycocotools.coco import COCO
+from pycocotools.cocoeval import COCOeval
+from pycocotools import mask as COCOmask
+
+class coco(imdb):
+  def __init__(self, image_set, year):
+    imdb.__init__(self, 'coco_' + year + '_' + image_set)
+    # COCO specific config options
+    self.config = {'use_salt': True,
+                   'cleanup': True}
+    # name, paths
+    self._year = year
+    self._image_set = image_set
+    self._data_path = osp.join(cfg.DATA_DIR, 'coco')
+    # load COCO API, classes, class <-> id mappings
+    self._COCO = COCO(self._get_ann_file())
+    cats = self._COCO.loadCats(self._COCO.getCatIds())
+    self._classes = tuple(['__background__'] + [c['name'] for c in cats])
+    self._class_to_ind = dict(list(zip(self.classes, list(range(self.num_classes)))))
+    self._class_to_coco_cat_id = dict(list(zip([c['name'] for c in cats],
+                                               self._COCO.getCatIds())))
+    self._image_index = self._load_image_set_index()
+    # Default to roidb handler
+    self.set_proposal_method('gt')
+    self.competition_mode(False)
+
+    # Some image sets are "views" (i.e. subsets) into others.
+    # For example, minival2014 is a random 5000 image subset of val2014.
+    # This mapping tells us where the view's images and proposals come from.
+    self._view_map = {
+      'minival2014': 'val2014',  # 5k val2014 subset
+      'valminusminival2014': 'val2014',  # val2014 \setminus minival2014
+      'test-dev2015': 'test2015',
+    }
+    coco_name = image_set + year  # e.g., "val2014"
+    self._data_name = (self._view_map[coco_name]
+                       if coco_name in self._view_map
+                       else coco_name)
+    # Dataset splits that have ground-truth annotations (test splits
+    # do not have gt annotations)
+    self._gt_splits = ('train', 'val', 'minival')
+
+  def _get_ann_file(self):
+    prefix = 'instances' if self._image_set.find('test') == -1 \
+      else 'image_info'
+    return osp.join(self._data_path, 'annotations',
+                    prefix + '_' + self._image_set + self._year + '.json')
+
+  def _load_image_set_index(self):
+    """
+    Load image ids.
+    """
+    image_ids = self._COCO.getImgIds()
+    return image_ids
+
+  def _get_widths(self):
+    anns = self._COCO.loadImgs(self._image_index)
+    widths = [ann['width'] for ann in anns]
+    return widths
+
+  def image_path_at(self, i):
+    """
+    Return the absolute path to image i in the image sequence.
+    """
+    return self.image_path_from_index(self._image_index[i])
+
+  def image_path_from_index(self, index):
+    """
+    Construct an image path from the image's "index" identifier.
+    """
+    # Example image path for index=119993:
+    #   images/train2014/COCO_train2014_000000119993.jpg
+    file_name = ('COCO_' + self._data_name + '_' +
+                 str(index).zfill(12) + '.jpg')
+    image_path = osp.join(self._data_path, 'images',
+                          self._data_name, file_name)
+    assert osp.exists(image_path), \
+      'Path does not exist: {}'.format(image_path)
+    return image_path
+
+  def gt_roidb(self):
+    """
+    Return the database of ground-truth regions of interest.
+    This function loads/saves from/to a cache file to speed up future calls.
+    """
+    cache_file = osp.join(self.cache_path, self.name + '_gt_roidb.pkl')
+    if osp.exists(cache_file):
+      with open(cache_file, 'rb') as fid:
+        roidb = pickle.load(fid)
+      print('{} gt roidb loaded from {}'.format(self.name, cache_file))
+      return roidb
+
+    gt_roidb = [self._load_coco_annotation(index)
+                for index in self._image_index]
+
+    with open(cache_file, 'wb') as fid:
+      pickle.dump(gt_roidb, fid, pickle.HIGHEST_PROTOCOL)
+    print('wrote gt roidb to {}'.format(cache_file))
+    return gt_roidb
+
+  def _load_coco_annotation(self, index):
+    """
+    Loads COCO bounding-box instance annotations. Crowd instances are
+    handled by marking their overlaps (with all categories) to -1. This
+    overlap value means that crowd "instances" are excluded from training.
+    """
+    im_ann = self._COCO.loadImgs(index)[0]
+    width = im_ann['width']
+    height = im_ann['height']
+
+    annIds = self._COCO.getAnnIds(imgIds=index, iscrowd=None)
+    objs = self._COCO.loadAnns(annIds)
+    # Sanitize bboxes -- some are invalid
+    valid_objs = []
+    for obj in objs:
+      x1 = np.max((0, obj['bbox'][0]))
+      y1 = np.max((0, obj['bbox'][1]))
+      x2 = np.min((width - 1, x1 + np.max((0, obj['bbox'][2] - 1))))
+      y2 = np.min((height - 1, y1 + np.max((0, obj['bbox'][3] - 1))))
+      if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
+        obj['clean_bbox'] = [x1, y1, x2, y2]
+        valid_objs.append(obj)
+    objs = valid_objs
+    num_objs = len(objs)
+
+    boxes = np.zeros((num_objs, 4), dtype=np.uint16)
+    gt_classes = np.zeros((num_objs), dtype=np.int32)
+    overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
+    seg_areas = np.zeros((num_objs), dtype=np.float32)
+
+    # Lookup table to map from COCO category ids to our internal class
+    # indices
+    coco_cat_id_to_class_ind = dict([(self._class_to_coco_cat_id[cls],
+                                      self._class_to_ind[cls])
+                                     for cls in self._classes[1:]])
+
+    for ix, obj in enumerate(objs):
+      cls = coco_cat_id_to_class_ind[obj['category_id']]
+      boxes[ix, :] = obj['clean_bbox']
+      gt_classes[ix] = cls
+      seg_areas[ix] = obj['area']
+      if obj['iscrowd']:
+        # Set overlap to -1 for all classes for crowd objects
+        # so they will be excluded during training
+        overlaps[ix, :] = -1.0
+      else:
+        overlaps[ix, cls] = 1.0
+
+    ds_utils.validate_boxes(boxes, width=width, height=height)
+    overlaps = scipy.sparse.csr_matrix(overlaps)
+    return {'width': width,
+            'height': height,
+            'boxes': boxes,
+            'gt_classes': gt_classes,
+            'gt_overlaps': overlaps,
+            'flipped': False,
+            'seg_areas': seg_areas}
+
+  def _get_widths(self):
+    return [r['width'] for r in self.roidb]
+
+  def append_flipped_images(self):
+    num_images = self.num_images
+    widths = self._get_widths()
+    for i in range(num_images):
+      boxes = self.roidb[i]['boxes'].copy()
+      oldx1 = boxes[:, 0].copy()
+      oldx2 = boxes[:, 2].copy()
+      boxes[:, 0] = widths[i] - oldx2 - 1
+      boxes[:, 2] = widths[i] - oldx1 - 1
+      assert (boxes[:, 2] >= boxes[:, 0]).all()
+      entry = {'width': widths[i],
+               'height': self.roidb[i]['height'],
+               'boxes': boxes,
+               'gt_classes': self.roidb[i]['gt_classes'],
+               'gt_overlaps': self.roidb[i]['gt_overlaps'],
+               'flipped': True,
+               'seg_areas': self.roidb[i]['seg_areas']}
+
+      self.roidb.append(entry)
+    self._image_index = self._image_index * 2
+
+  def _get_box_file(self, index):
+    # first 14 chars / first 22 chars / all chars + .mat
+    # COCO_val2014_0/COCO_val2014_000000447/COCO_val2014_000000447991.mat
+    file_name = ('COCO_' + self._data_name +
+                 '_' + str(index).zfill(12) + '.mat')
+    return osp.join(file_name[:14], file_name[:22], file_name)
+
+  def _print_detection_eval_metrics(self, coco_eval):
+    IoU_lo_thresh = 0.5
+    IoU_hi_thresh = 0.95
+
+    def _get_thr_ind(coco_eval, thr):
+      ind = np.where((coco_eval.params.iouThrs > thr - 1e-5) &
+                     (coco_eval.params.iouThrs < thr + 1e-5))[0][0]
+      iou_thr = coco_eval.params.iouThrs[ind]
+      assert np.isclose(iou_thr, thr)
+      return ind
+
+    ind_lo = _get_thr_ind(coco_eval, IoU_lo_thresh)
+    ind_hi = _get_thr_ind(coco_eval, IoU_hi_thresh)
+    # precision has dims (iou, recall, cls, area range, max dets)
+    # area range index 0: all area ranges
+    # max dets index 2: 100 per image
+    precision = \
+      coco_eval.eval['precision'][ind_lo:(ind_hi + 1), :, :, 0, 2]
+    ap_default = np.mean(precision[precision > -1])
+    print(('~~~~ Mean and per-category AP @ IoU=[{:.2f},{:.2f}] '
+           '~~~~').format(IoU_lo_thresh, IoU_hi_thresh))
+    print('{:.1f}'.format(100 * ap_default))
+    for cls_ind, cls in enumerate(self.classes):
+      if cls == '__background__':
+        continue
+      # minus 1 because of __background__
+      precision = coco_eval.eval['precision'][ind_lo:(ind_hi + 1), :, cls_ind - 1, 0, 2]
+      ap = np.mean(precision[precision > -1])
+      print('{:.1f}'.format(100 * ap))
+
+    print('~~~~ Summary metrics ~~~~')
+    coco_eval.summarize()
+
+  def _do_detection_eval(self, res_file, output_dir):
+    ann_type = 'bbox'
+    coco_dt = self._COCO.loadRes(res_file)
+    coco_eval = COCOeval(self._COCO, coco_dt)
+    coco_eval.params.useSegm = (ann_type == 'segm')
+    coco_eval.evaluate()
+    coco_eval.accumulate()
+    self._print_detection_eval_metrics(coco_eval)
+    eval_file = osp.join(output_dir, 'detection_results.pkl')
+    with open(eval_file, 'wb') as fid:
+      pickle.dump(coco_eval, fid, pickle.HIGHEST_PROTOCOL)
+    print('Wrote COCO eval results to: {}'.format(eval_file))
+
+  def _coco_results_one_category(self, boxes, cat_id):
+    results = []
+    for im_ind, index in enumerate(self.image_index):
+      dets = boxes[im_ind].astype(np.float)
+      if dets == []:
+        continue
+      scores = dets[:, -1]
+      xs = dets[:, 0]
+      ys = dets[:, 1]
+      ws = dets[:, 2] - xs + 1
+      hs = dets[:, 3] - ys + 1
+      results.extend(
+        [{'image_id': index,
+          'category_id': cat_id,
+          'bbox': [xs[k], ys[k], ws[k], hs[k]],
+          'score': scores[k]} for k in range(dets.shape[0])])
+    return results
+
+  def _write_coco_results_file(self, all_boxes, res_file):
+    # [{"image_id": 42,
+    #   "category_id": 18,
+    #   "bbox": [258.15,41.29,348.26,243.78],
+    #   "score": 0.236}, ...]
+    results = []
+    for cls_ind, cls in enumerate(self.classes):
+      if cls == '__background__':
+        continue
+      print('Collecting {} results ({:d}/{:d})'.format(cls, cls_ind,
+                                                       self.num_classes - 1))
+      coco_cat_id = self._class_to_coco_cat_id[cls]
+      results.extend(self._coco_results_one_category(all_boxes[cls_ind],
+                                                     coco_cat_id))
+    print('Writing results json to {}'.format(res_file))
+    with open(res_file, 'w') as fid:
+      json.dump(results, fid)
+
+  def evaluate_detections(self, all_boxes, output_dir):
+    res_file = osp.join(output_dir, ('detections_' +
+                                     self._image_set +
+                                     self._year +
+                                     '_results'))
+    if self.config['use_salt']:
+      res_file += '_{}'.format(str(uuid.uuid4()))
+    res_file += '.json'
+    self._write_coco_results_file(all_boxes, res_file)
+    # Only do evaluation on non-test sets
+    if self._image_set.find('test') == -1:
+      self._do_detection_eval(res_file, output_dir)
+    # Optionally cleanup results json file
+    if self.config['cleanup']:
+      os.remove(res_file)
+
+  def competition_mode(self, on):
+    if on:
+      self.config['use_salt'] = False
+      self.config['cleanup'] = False
+    else:
+      self.config['use_salt'] = True
+      self.config['cleanup'] = True
diff --git a/lib/datasets/ds_utils.py b/lib/datasets/ds_utils.py
new file mode 100644
index 0000000..fd5ca4b
--- /dev/null
+++ b/lib/datasets/ds_utils.py
@@ -0,0 +1,49 @@
+# --------------------------------------------------------
+# Fast/er R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+
+def unique_boxes(boxes, scale=1.0):
+  """Return indices of unique boxes."""
+  v = np.array([1, 1e3, 1e6, 1e9])
+  hashes = np.round(boxes * scale).dot(v)
+  _, index = np.unique(hashes, return_index=True)
+  return np.sort(index)
+
+
+def xywh_to_xyxy(boxes):
+  """Convert [x y w h] box format to [x1 y1 x2 y2] format."""
+  return np.hstack((boxes[:, 0:2], boxes[:, 0:2] + boxes[:, 2:4] - 1))
+
+
+def xyxy_to_xywh(boxes):
+  """Convert [x1 y1 x2 y2] box format to [x y w h] format."""
+  return np.hstack((boxes[:, 0:2], boxes[:, 2:4] - boxes[:, 0:2] + 1))
+
+
+def validate_boxes(boxes, width=0, height=0):
+  """Check that a set of boxes are valid."""
+  x1 = boxes[:, 0]
+  y1 = boxes[:, 1]
+  x2 = boxes[:, 2]
+  y2 = boxes[:, 3]
+  assert (x1 >= 0).all()
+  assert (y1 >= 0).all()
+  assert (x2 >= x1).all()
+  assert (y2 >= y1).all()
+  assert (x2 < width).all()
+  assert (y2 < height).all()
+
+
+def filter_small_boxes(boxes, min_size):
+  w = boxes[:, 2] - boxes[:, 0]
+  h = boxes[:, 3] - boxes[:, 1]
+  keep = np.where((w >= min_size) & (h > min_size))[0]
+  return keep
diff --git a/lib/datasets/factory.py b/lib/datasets/factory.py
new file mode 100644
index 0000000..c1f823b
--- /dev/null
+++ b/lib/datasets/factory.py
@@ -0,0 +1,57 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+"""Factory method for easily getting imdbs by name."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+__sets = {}
+from datasets.pascal_voc import pascal_voc
+
+#from datasets.coco import coco
+
+#from datasets.coco import coco
+
+import numpy as np
+
+# Set up voc_<year>_<split> 
+for year in ['2007', '2012']:
+  for split in ['train', 'val', 'trainval', 'test']:
+    name = 'voc_{}_{}'.format(year, split)
+    __sets[name] = (lambda split=split, year=year: pascal_voc(split, year))
+
+for year in ['2007', '2012']:
+  for split in ['train', 'val', 'trainval', 'test']:
+    name = 'voc_{}_{}_diff'.format(year, split)
+    __sets[name] = (lambda split=split, year=year: pascal_voc(split, year, use_diff=True))
+
+
+# Set up coco_2014_<split>
+for year in ['2014']:
+  for split in ['train', 'val', 'minival', 'valminusminival', 'trainval']:
+    name = 'coco_{}_{}'.format(year, split)
+    __sets[name] = (lambda split=split, year=year: coco(split, year))
+
+# Set up coco_2015_<split>
+for year in ['2015']:
+  for split in ['test', 'test-dev']:
+    name = 'coco_{}_{}'.format(year, split)
+    __sets[name] = (lambda split=split, year=year: coco(split, year))
+
+
+
+def get_imdb(name):
+  """Get an imdb (image database) by name."""
+  if name not in __sets:
+    raise KeyError('Unknown dataset: {}'.format(name))
+  return __sets[name]()
+
+
+def list_imdbs():
+  """List all registered imdbs."""
+  return list(__sets.keys())
diff --git a/lib/datasets/imdb.py b/lib/datasets/imdb.py
new file mode 100644
index 0000000..03cea0a
--- /dev/null
+++ b/lib/datasets/imdb.py
@@ -0,0 +1,264 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick and Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import os.path as osp
+import PIL
+from utils.bbox import bbox_overlaps
+import numpy as np
+import scipy.sparse
+from model.config import cfg
+
+
+class imdb(object):
+  """Image database."""
+
+  def __init__(self, name, classes=None):
+    self._name = name
+    self._num_classes = 0
+    if not classes:
+      self._classes = []
+    else:
+      self._classes = classes
+    self._image_index = []
+    self._obj_proposer = 'selective_search'
+    self._roidb = None
+    self._roidb_handler = self.default_roidb
+    # Use this dict for storing dataset specific config options
+    self.config = {}
+
+  @property
+  def name(self):
+    return self._name
+
+  @property
+  def num_classes(self):
+    return len(self._classes)
+
+  @property
+  def classes(self):
+    return self._classes
+
+  @property
+  def image_index(self):
+    return self._image_index
+
+  @property
+  def roidb_handler(self):
+    return self._roidb_handler
+
+  @roidb_handler.setter
+  def roidb_handler(self, val):
+    self._roidb_handler = val
+
+  def set_proposal_method(self, method):
+    method = eval('self.' + method + '_roidb')
+    self.roidb_handler = method
+
+  @property
+  def roidb(self):
+    # A roidb is a list of dictionaries, each with the following keys:
+    #   boxes
+    #   gt_overlaps
+    #   gt_classes
+    #   flipped
+    if self._roidb is not None:
+      return self._roidb
+    self._roidb = self.roidb_handler()
+    return self._roidb
+
+  @property
+  def cache_path(self):
+    cache_path = osp.abspath(osp.join(cfg.DATA_DIR, 'cache'))
+    if not os.path.exists(cache_path):
+      os.makedirs(cache_path)
+    return cache_path
+
+  @property
+  def num_images(self):
+    return len(self.image_index)
+
+  def image_path_at(self, i):
+    raise NotImplementedError
+
+  def default_roidb(self):
+    raise NotImplementedError
+
+  def evaluate_detections(self, all_boxes, output_dir=None):
+    """
+    all_boxes is a list of length number-of-classes.
+    Each list element is a list of length number-of-images.
+    Each of those list elements is either an empty list []
+    or a numpy array of detection.
+
+    all_boxes[class][image] = [] or np.array of shape #dets x 5
+    """
+    raise NotImplementedError
+
+  def _get_widths(self):
+    return [PIL.Image.open(self.image_path_at(i)).size[0]
+            for i in range(self.num_images)]
+
+  def append_flipped_images(self):
+    num_images = self.num_images
+    widths = self._get_widths()
+    for i in range(num_images):
+      boxes = self.roidb[i]['boxes'].copy()
+      oldx1 = boxes[:, 0].copy()
+      oldx2 = boxes[:, 2].copy()
+      boxes[:, 0] = widths[i] - oldx2 - 1
+      boxes[:, 2] = widths[i] - oldx1 - 1
+      assert (boxes[:, 2] >= boxes[:, 0]).all()
+      #print(self.roidb[i].keys())
+      entry = {'boxes': boxes,
+               'gt_overlaps': self.roidb[i]['gt_overlaps'],
+               'gt_classes': self.roidb[i]['gt_classes'],
+               'flipped': True,
+               'image_level_labels':self.roidb[i]['image_level_labels']}
+      self.roidb.append(entry)
+    self._image_index = self._image_index * 2
+
+  def evaluate_recall(self, candidate_boxes=None, thresholds=None,
+                      area='all', limit=None):
+    """Evaluate detection proposal recall metrics.
+
+    Returns:
+        results: dictionary of results with keys
+            'ar': average recall
+            'recalls': vector recalls at each IoU overlap threshold
+            'thresholds': vector of IoU overlap thresholds
+            'gt_overlaps': vector of all ground-truth overlaps
+    """
+    # Record max overlap value for each gt box
+    # Return vector of overlap values
+    areas = {'all': 0, 'small': 1, 'medium': 2, 'large': 3,
+             '96-128': 4, '128-256': 5, '256-512': 6, '512-inf': 7}
+    area_ranges = [[0 ** 2, 1e5 ** 2],  # all
+                   [0 ** 2, 32 ** 2],  # small
+                   [32 ** 2, 96 ** 2],  # medium
+                   [96 ** 2, 1e5 ** 2],  # large
+                   [96 ** 2, 128 ** 2],  # 96-128
+                   [128 ** 2, 256 ** 2],  # 128-256
+                   [256 ** 2, 512 ** 2],  # 256-512
+                   [512 ** 2, 1e5 ** 2],  # 512-inf
+                   ]
+    assert area in areas, 'unknown area range: {}'.format(area)
+    area_range = area_ranges[areas[area]]
+    gt_overlaps = np.zeros(0)
+    num_pos = 0
+    for i in range(self.num_images):
+      # Checking for max_overlaps == 1 avoids including crowd annotations
+      # (...pretty hacking :/)
+      max_gt_overlaps = self.roidb[i]['gt_overlaps'].toarray().max(axis=1)
+      gt_inds = np.where((self.roidb[i]['gt_classes'] > 0) &
+                         (max_gt_overlaps == 1))[0]
+      gt_boxes = self.roidb[i]['boxes'][gt_inds, :]
+      gt_areas = self.roidb[i]['seg_areas'][gt_inds]
+      valid_gt_inds = np.where((gt_areas >= area_range[0]) &
+                               (gt_areas <= area_range[1]))[0]
+      gt_boxes = gt_boxes[valid_gt_inds, :]
+      num_pos += len(valid_gt_inds)
+
+      if candidate_boxes is None:
+        # If candidate_boxes is not supplied, the default is to use the
+        # non-ground-truth boxes from this roidb
+        non_gt_inds = np.where(self.roidb[i]['gt_classes'] == 0)[0]
+        boxes = self.roidb[i]['boxes'][non_gt_inds, :]
+      else:
+        boxes = candidate_boxes[i]
+      if boxes.shape[0] == 0:
+        continue
+      if limit is not None and boxes.shape[0] > limit:
+        boxes = boxes[:limit, :]
+
+      overlaps = bbox_overlaps(boxes.astype(np.float),
+                               gt_boxes.astype(np.float))
+
+      _gt_overlaps = np.zeros((gt_boxes.shape[0]))
+      for j in range(gt_boxes.shape[0]):
+        # find which proposal box maximally covers each gt box
+        argmax_overlaps = overlaps.argmax(axis=0)
+        # and get the iou amount of coverage for each gt box
+        max_overlaps = overlaps.max(axis=0)
+        # find which gt box is 'best' covered (i.e. 'best' = most iou)
+        gt_ind = max_overlaps.argmax()
+        gt_ovr = max_overlaps.max()
+        assert (gt_ovr >= 0)
+        # find the proposal box that covers the best covered gt box
+        box_ind = argmax_overlaps[gt_ind]
+        # record the iou coverage of this gt box
+        _gt_overlaps[j] = overlaps[box_ind, gt_ind]
+        assert (_gt_overlaps[j] == gt_ovr)
+        # mark the proposal box and the gt box as used
+        overlaps[box_ind, :] = -1
+        overlaps[:, gt_ind] = -1
+      # append recorded iou coverage level
+      gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
+
+    gt_overlaps = np.sort(gt_overlaps)
+    if thresholds is None:
+      step = 0.05
+      thresholds = np.arange(0.5, 0.95 + 1e-5, step)
+    recalls = np.zeros_like(thresholds)
+    # compute recall for each iou threshold
+    for i, t in enumerate(thresholds):
+      recalls[i] = (gt_overlaps >= t).sum() / float(num_pos)
+    # ar = 2 * np.trapz(recalls, thresholds)
+    ar = recalls.mean()
+    return {'ar': ar, 'recalls': recalls, 'thresholds': thresholds,
+            'gt_overlaps': gt_overlaps}
+
+  def create_roidb_from_box_list(self, box_list, gt_roidb):
+    assert len(box_list) == self.num_images, \
+      'Number of boxes must match number of ground-truth images'
+    roidb = []
+    for i in range(self.num_images):
+      boxes = box_list[i]
+      num_boxes = boxes.shape[0]
+      overlaps = np.zeros((num_boxes, self.num_classes), dtype=np.float32)
+
+      if gt_roidb is not None and gt_roidb[i]['boxes'].size > 0:
+        gt_boxes = gt_roidb[i]['boxes']
+        gt_classes = gt_roidb[i]['gt_classes']
+        gt_overlaps = bbox_overlaps(boxes.astype(np.float),
+                                    gt_boxes.astype(np.float))
+        argmaxes = gt_overlaps.argmax(axis=1)
+        maxes = gt_overlaps.max(axis=1)
+        I = np.where(maxes > 0)[0]
+        overlaps[I, gt_classes[argmaxes[I]]] = maxes[I]
+
+      overlaps = scipy.sparse.csr_matrix(overlaps)
+      roidb.append({
+        'boxes': boxes,
+        'gt_classes': -np.ones((num_boxes,), dtype=np.int32),
+        'gt_overlaps': overlaps,
+        'flipped': False,
+        'seg_areas': np.zeros((num_boxes,), dtype=np.float32),
+        'image_level_labels':np.zeros((1, self.num_classes), dtype=np.int32)
+      })
+    return roidb
+
+  @staticmethod
+  def merge_roidbs(a, b):
+    assert len(a) == len(b)
+    for i in range(len(a)):
+      a[i]['boxes'] = np.vstack((a[i]['boxes'], b[i]['boxes']))
+      a[i]['gt_classes'] = np.hstack((a[i]['gt_classes'],
+                                      b[i]['gt_classes']))
+      a[i]['gt_overlaps'] = scipy.sparse.vstack([a[i]['gt_overlaps'],
+                                                 b[i]['gt_overlaps']])
+      a[i]['seg_areas'] = np.hstack((a[i]['seg_areas'],
+                                     b[i]['seg_areas']))
+      a[i]['image_level_labels'] = a[i]['image_level_labels']
+    return a
+
+  def competition_mode(self, on):
+    """Turn competition mode on or off."""
+    pass
diff --git a/lib/datasets/pascal_voc.py b/lib/datasets/pascal_voc.py
new file mode 100644
index 0000000..6bed8f5
--- /dev/null
+++ b/lib/datasets/pascal_voc.py
@@ -0,0 +1,417 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick and Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+from datasets.imdb import imdb
+import datasets.ds_utils as ds_utils
+import xml.etree.ElementTree as ET
+import numpy as np
+import scipy.sparse
+import scipy.io as sio
+import pickle
+import subprocess
+import uuid
+from .voc_eval import voc_eval
+from model.config import cfg
+
+
+class pascal_voc(imdb):
+  def __init__(self, image_set, year, use_diff=False):
+    name = 'voc_' + year + '_' + image_set
+    if use_diff:
+      name += '_diff'
+    imdb.__init__(self, name)
+    self._year = year
+    self._image_set = image_set
+    self._devkit_path = self._get_default_path()
+    self._data_path = os.path.join(self._devkit_path, 'VOC' + self._year)
+    self._classes = ('aeroplane', 'bicycle', 'bird', 'boat',
+                     'bottle', 'bus', 'car', 'cat', 'chair',
+                     'cow', 'diningtable', 'dog', 'horse',
+                     'motorbike', 'person', 'pottedplant',
+                     'sheep', 'sofa', 'train', 'tvmonitor')
+    self._class_to_ind = dict(list(zip(self.classes, list(range(self.num_classes)))))
+    self._image_ext = '.jpg'
+    self._image_index = self._load_image_set_index()
+    # Default to roidb handler
+    self._roidb_handler = self.selective_search_roidb
+    #self._roidb_handler = self.gt_roidb
+    self._salt = str(uuid.uuid4())
+    self._comp_id = 'comp4'
+
+    # PASCAL specific config options
+    self.config = {'cleanup': True,
+                   'use_salt': True,
+                   'use_diff': use_diff,
+                   'matlab_eval': False,
+                   'rpn_file': None,
+                   'min_size': 20}
+
+    assert os.path.exists(self._devkit_path), \
+      'VOCdevkit path does not exist: {}'.format(self._devkit_path)
+    assert os.path.exists(self._data_path), \
+      'Path does not exist: {}'.format(self._data_path)
+
+  def image_path_at(self, i):
+    """
+    Return the absolute path to image i in the image sequence.
+    """
+    return self.image_path_from_index(self._image_index[i])
+
+  def image_path_from_index(self, index):
+    """
+    Construct an image path from the image's "index" identifier.
+    """
+    image_path = os.path.join(self._data_path, 'JPEGImages',
+                              index + self._image_ext)
+    assert os.path.exists(image_path), \
+      'Path does not exist: {}'.format(image_path)
+    return image_path
+
+  def _load_image_set_index(self):
+    """
+    Load the indexes listed in this dataset's image set file.
+    """
+    # Example path to image set file:
+    # self._devkit_path + /VOCdevkit2007/VOC2007/ImageSets/Main/val.txt
+    image_set_file = os.path.join(self._data_path, 'ImageSets', 'Main',
+                                  self._image_set + '.txt')
+    assert os.path.exists(image_set_file), \
+      'Path does not exist: {}'.format(image_set_file)
+    with open(image_set_file) as f:
+      image_index = [x.strip() for x in f.readlines()]
+    return image_index
+
+  def _get_default_path(self):
+    """
+    Return the default path where PASCAL VOC is expected to be installed.
+    """
+    return os.path.join(cfg.DATA_DIR, 'VOCdevkit' + self._year)
+
+  def gt_roidb(self):
+    """
+    Return the database of ground-truth regions of interest.
+
+    This function loads/saves from/to a cache file to speed up future calls.
+    """
+    cache_file = os.path.join(self.cache_path, self.name + '_gt_roidb.pkl')
+    if os.path.exists(cache_file):
+      with open(cache_file, 'rb') as fid:
+        try:
+          roidb = pickle.load(fid)
+        except:
+          roidb = pickle.load(fid, encoding='bytes')
+      print('{} gt roidb loaded from {}'.format(self.name, cache_file))
+      return roidb
+
+    #gt_roidb = [self._load_pascal_annotation_retrain(index)
+    gt_roidb = [self._load_pascal_annotation(index)
+                for index in self.image_index]
+    with open(cache_file, 'wb') as fid:
+      pickle.dump(gt_roidb, fid, pickle.HIGHEST_PROTOCOL)
+    print('wrote gt roidb to {}'.format(cache_file))
+
+    return gt_roidb
+
+  def selective_search_roidb(self):
+    """
+    Return the database of selective search regions of interest.
+    Ground-truth ROIs are also included.
+
+    This function loads/saves from/to a cache file to speed up future calls.
+    """
+    cache_file = os.path.join(self.cache_path,
+                              self.name + '_selective_search_roidb.pkl')
+
+    if os.path.exists(cache_file):
+        with open(cache_file, 'rb') as fid:
+            roidb = pickle.load(fid)
+        print('{} ss roidb loaded from {}'.format(self.name, cache_file))
+
+        return roidb
+
+    if int(self._year) == 2007 or self._image_set != 'test':
+        gt_roidb = self.gt_roidb()
+        ss_roidb = self._load_selective_search_roidb(gt_roidb)
+        roidb = imdb.merge_roidbs(gt_roidb, ss_roidb)
+    else:
+        roidb = self._load_selective_search_roidb(None)
+    with open(cache_file, 'wb') as fid:
+        pickle.dump(roidb, fid, pickle.HIGHEST_PROTOCOL)
+    print('wrote ss roidb to {}'.format(cache_file))
+
+    return roidb
+
+  def rpn_roidb(self):
+    if int(self._year) == 2007 or self._image_set != 'test':
+      gt_roidb = self.gt_roidb()
+      rpn_roidb = self._load_rpn_roidb(gt_roidb)
+      roidb = imdb.merge_roidbs(gt_roidb, rpn_roidb)
+    else:
+      roidb = self._load_rpn_roidb(None)
+
+    return roidb
+
+  def _load_rpn_roidb(self, gt_roidb):
+    filename = self.config['rpn_file']
+    print('loading {}'.format(filename))
+    assert os.path.exists(filename), \
+      'rpn data not found at: {}'.format(filename)
+    with open(filename, 'rb') as f:
+      box_list = pickle.load(f)
+    return self.create_roidb_from_box_list(box_list, gt_roidb)
+
+  def _load_selective_search_roidb(self, gt_roidb):
+    filename = os.path.abspath(os.path.join(cfg.DATA_DIR,
+                                            'selective_search_data',
+                                            self.name + '.mat'))
+    assert os.path.exists(filename), \
+        'Selective search data not found at: {}'.format(filename)
+    raw_data = sio.loadmat(filename)['boxes'].ravel()
+
+    box_list = []
+    for i in range(raw_data.shape[0]):
+        boxes = raw_data[i][:, (1, 0, 3, 2)] - 1
+        keep = ds_utils.unique_boxes(boxes)
+        boxes = boxes[keep, :]
+        keep = ds_utils.filter_small_boxes(boxes, self.config['min_size'])
+        boxes = boxes[keep, :]
+        box_list.append(boxes)
+
+    return self.create_roidb_from_box_list(box_list, gt_roidb)
+
+
+  def _load_pascal_annotation(self, index):
+    """
+    Load image and bounding boxes info from XML file in the PASCAL VOC
+    format.
+    """
+    filename = os.path.join(self._data_path, 'Annotations', index + '.xml')
+    tree = ET.parse(filename)
+    objs = tree.findall('object')
+    if not self.config['use_diff']:
+      # Exclude the samples labeled as difficult
+      non_diff_objs = [
+        obj for obj in objs if int(obj.find('difficult').text) == 0]
+      # if len(non_diff_objs) != len(objs):
+      #     print 'Removed {} difficult objects'.format(
+      #         len(objs) - len(non_diff_objs))
+      objs = non_diff_objs
+    num_objs = len(objs)
+
+    boxes = np.zeros((num_objs, 4), dtype=np.uint16)
+    gt_classes = np.zeros((num_objs), dtype=np.int32)
+    overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
+    # "Seg" area for pascal is just the box area
+    seg_areas = np.zeros((num_objs), dtype=np.float32)
+    
+    image_level_labels = -np.ones((1, self.num_classes), dtype=np.int32)
+    
+    # Load object bounding boxes into a data frame.
+    for ix, obj in enumerate(objs):
+      bbox = obj.find('bndbox')
+      # Make pixel indexes 0-based
+      x1 = float(bbox.find('xmin').text) - 1
+      y1 = float(bbox.find('ymin').text) - 1
+      x2 = float(bbox.find('xmax').text) - 1
+      y2 = float(bbox.find('ymax').text) - 1
+      cls = self._class_to_ind[obj.find('name').text.lower().strip()]
+      image_level_labels[0][cls] = 1
+      boxes[ix, :] = [x1, y1, x2, y2]
+      gt_classes[ix] = cls
+      overlaps[ix, cls] = 1.0
+      seg_areas[ix] = (x2 - x1 + 1) * (y2 - y1 + 1)
+
+    overlaps = scipy.sparse.csr_matrix(overlaps)
+
+    return {'boxes': boxes,
+            'gt_classes': gt_classes,
+            'gt_overlaps': overlaps,
+            'flipped': False,
+            'seg_areas': seg_areas,
+            'image_level_labels':image_level_labels}
+
+  def _load_pascal_annotation_retrain(self, index):
+    """
+    Load image and bounding boxes info from XML file in the PASCAL VOC
+    format./home/wan/Servers/G3/Cloud/G5/MELM-VGG16/output/VOC2007/VGG16/PsuedoGT/Top1
+    """
+    # TopThresh0.7   Top1
+    #gt_folder = '/root/Cloud/G5/MELM-VGG16/output/VOC2007/VGG16/PsuedoGT/Top1/'
+    gt_folder = '/root/Cloud/G5/MELM-VGG16/output/VOC2007/VGG16/PsuedoGT/Detections/'
+    #gt_folder = '/root/Cloud/G5/MELM-VGG16/output/VOC2007/VGG16/PsuedoGT/Top10/'
+    print('Loading ', gt_folder, ': ', index)
+    filename = os.path.join(gt_folder, index + '.txt')
+    objs = np.loadtxt(filename)
+    if objs.ndim == 1:
+      num_objs = 1
+      objs = objs.reshape(1,-1)
+    else:
+      num_objs = objs.shape[0]
+
+    boxes = np.zeros((num_objs, 4), dtype=np.uint16)
+    gt_classes = np.zeros((num_objs), dtype=np.int32)
+    overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
+    seg_areas = np.zeros((num_objs), dtype=np.float32)
+
+    # Load object bounding boxes into a data frame.
+    for i in range(num_objs):
+      # Make pixel indexes 0-based
+      x1 = objs[i][1]
+      y1 = objs[i][2]
+      x2 = objs[i][3]
+      y2 = objs[i][4]
+      boxes[i, :] = [x1, y1, x2, y2]
+      gt_classes[i] = int(objs[i][0])
+      overlaps[i, int(objs[i][0])] = 1.0
+      seg_areas[i] = (x2 - x1 + 1) * (y2 - y1 + 1)
+
+    overlaps = scipy.sparse.csr_matrix(overlaps)
+
+    return {'boxes' : boxes,
+            'gt_classes': gt_classes,
+            'gt_overlaps' : overlaps,
+            'flipped' : False,
+            'seg_areas': seg_areas}
+
+
+  def _get_comp_id(self):
+    comp_id = (self._comp_id + '_' + self._salt if self.config['use_salt']
+               else self._comp_id)
+    return comp_id
+
+  def _get_voc_results_file_template(self):
+    # VOCdevkit/results/VOC2007/Main/<comp_id>_det_test_aeroplane.txt
+    filename = self._get_comp_id() + '_det_' + self._image_set + '_{:s}.txt'
+    path = os.path.join(
+      self._devkit_path,
+      'results',
+      'VOC' + self._year,
+      'Main',
+      filename)
+    return path
+
+  def _write_voc_results_file(self, all_boxes):
+    for cls_ind, cls in enumerate(self.classes):
+      if cls == '__background__':
+        continue
+      print('Writing {} VOC results file'.format(cls))
+      filename = self._get_voc_results_file_template().format(cls)
+
+      new_filename = os.path.join(self._devkit_path, 'results', 'VOC' + self._year, 'Main')  # ------------------------
+
+      if not os.path.exists(new_filename):
+        os.makedirs(new_filename)
+      with open(filename, 'wt') as f:
+        for im_ind, index in enumerate(self.image_index):
+          dets = all_boxes[cls_ind][im_ind]
+          if dets == []:
+
+          # dets = all_boxes[cls_ind][im_ind].numpy()
+          # print(dets.shape)
+          # if dets.shape[0]:
+
+            continue
+          # the VOCdevkit expects 1-based indices
+          for k in range(dets.shape[0]):
+            f.write('{:s} {:.3f} {:.1f} {:.1f} {:.1f} {:.1f}\n'.
+                    format(index, dets[k, -1],
+                           dets[k, 0] + 1, dets[k, 1] + 1,
+                           dets[k, 2] + 1, dets[k, 3] + 1))
+
+  def _do_python_eval(self, output_dir='output'):
+    annopath = os.path.join(
+      self._devkit_path,
+      'VOC' + self._year,
+      'Annotations',
+      '{:s}.xml')
+    imagesetfile = os.path.join(
+      self._devkit_path,
+      'VOC' + self._year,
+      'ImageSets',
+      'Main',
+      self._image_set + '.txt')
+    cachedir = os.path.join(self._devkit_path, 'annotations_cache')
+    aps = []
+    # The PASCAL VOC metric changed in 2010
+    use_07_metric = True if int(self._year) < 2010 else False
+    print('VOC07 metric? ' + ('Yes' if use_07_metric else 'No'))
+    if not os.path.isdir(output_dir):
+      os.mkdir(output_dir)
+    for i, cls in enumerate(self._classes):
+      if cls == '__background__':
+        continue
+      filename = self._get_voc_results_file_template().format(cls)
+      rec, prec, ap = voc_eval(
+        filename, annopath, imagesetfile, cls, cachedir, ovthresh=0.5,
+        use_07_metric=use_07_metric, use_diff=self.config['use_diff'])
+      aps += [ap]
+      print(('AP for {} = {:.4f}'.format(cls, ap)))
+      with open(os.path.join(output_dir, cls + '_pr.pkl'), 'wb') as f:
+        pickle.dump({'rec': rec, 'prec': prec, 'ap': ap}, f)
+    print(('Mean AP = {:.4f}'.format(np.mean(aps))))
+    print('~~~~~~~~')
+    print('Results:')
+    for ap in aps:
+      print(('{:.3f}'.format(ap)))
+    print(('{:.3f}'.format(np.mean(aps))))
+    print('~~~~~~~~')
+    print('')
+    print('--------------------------------------------------------------')
+    print('Results computed with the **unofficial** Python eval code.')
+    print('Results should be very close to the official MATLAB eval code.')
+    print('Recompute with `./tools/reval.py --matlab ...` for your paper.')
+    print('-- Thanks, The Management')
+    print('--------------------------------------------------------------')
+
+  def _do_matlab_eval(self, output_dir='output'):
+    print('-----------------------------------------------------')
+    print('Computing results with the official MATLAB eval code.')
+    print('-----------------------------------------------------')
+    path = os.path.join(cfg.ROOT_DIR, 'lib', 'datasets',
+                        'VOCdevkit-matlab-wrapper')
+    cmd = 'cd {} && '.format(path)
+    cmd += '{:s} -nodisplay -nodesktop '.format(cfg.MATLAB)
+    cmd += '-r "dbstop if error; '
+    cmd += 'voc_eval(\'{:s}\',\'{:s}\',\'{:s}\',\'{:s}\'); quit;"' \
+      .format(self._devkit_path, self._get_comp_id(),
+              self._image_set, output_dir)
+    print(('Running:\n{}'.format(cmd)))
+    status = subprocess.call(cmd, shell=True)
+
+  def evaluate_detections(self, all_boxes, output_dir):
+    self._write_voc_results_file(all_boxes)
+    self._do_python_eval(output_dir)
+    if self.config['matlab_eval']:
+      self._do_matlab_eval(output_dir)
+    if self.config['cleanup']:
+      for cls in self._classes:
+        if cls == '__background__':
+          continue
+        filename = self._get_voc_results_file_template().format(cls)
+        os.remove(filename)
+
+  def competition_mode(self, on):
+    if on:
+      self.config['use_salt'] = False
+      self.config['cleanup'] = False
+    else:
+      self.config['use_salt'] = True
+      self.config['cleanup'] = True
+
+
+if __name__ == '__main__':
+  from datasets.pascal_voc import pascal_voc
+
+  d = pascal_voc('trainval', '2007')
+  res = d.roidb
+  from IPython import embed;
+
+  embed()
diff --git a/lib/datasets/tools/mcg_munge.py b/lib/datasets/tools/mcg_munge.py
new file mode 100644
index 0000000..1392aa3
--- /dev/null
+++ b/lib/datasets/tools/mcg_munge.py
@@ -0,0 +1,38 @@
+import os
+import sys
+
+"""Hacky tool to convert file system layout of MCG boxes downloaded from
+http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/mcg/
+so that it's consistent with those computed by Jan Hosang (see:
+http://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-
+  computing/research/object-recognition-and-scene-understanding/how-
+  good-are-detection-proposals-really/)
+
+NB: Boxes from the MCG website are in (y1, x1, y2, x2) order.
+Boxes from Hosang et al. are in (x1, y1, x2, y2) order.
+"""
+
+def munge(src_dir):
+    # stored as: ./MCG-COCO-val2014-boxes/COCO_val2014_000000193401.mat
+    # want:      ./MCG/mat/COCO_val2014_0/COCO_val2014_000000141/COCO_val2014_000000141334.mat
+
+    files = os.listdir(src_dir)
+    for fn in files:
+        base, ext = os.path.splitext(fn)
+        # first 14 chars / first 22 chars / all chars + .mat
+        # COCO_val2014_0/COCO_val2014_000000447/COCO_val2014_000000447991.mat
+        first = base[:14]
+        second = base[:22]
+        dst_dir = os.path.join('MCG', 'mat', first, second)
+        if not os.path.exists(dst_dir):
+            os.makedirs(dst_dir)
+        src = os.path.join(src_dir, fn)
+        dst = os.path.join(dst_dir, fn)
+        print 'MV: {} -> {}'.format(src, dst)
+        os.rename(src, dst)
+
+if __name__ == '__main__':
+    # src_dir should look something like:
+    #  src_dir = 'MCG-COCO-val2014-boxes'
+    src_dir = sys.argv[1]
+    munge(src_dir)
diff --git a/lib/datasets/voc_eval.py b/lib/datasets/voc_eval.py
new file mode 100644
index 0000000..17a9ee8
--- /dev/null
+++ b/lib/datasets/voc_eval.py
@@ -0,0 +1,214 @@
+# --------------------------------------------------------
+# Fast/er R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Bharath Hariharan
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import xml.etree.ElementTree as ET
+import os
+import pickle
+import numpy as np
+
+def parse_rec(filename):
+  """ Parse a PASCAL VOC xml file """
+  tree = ET.parse(filename)
+  objects = []
+  for obj in tree.findall('object'):
+    obj_struct = {}
+    obj_struct['name'] = obj.find('name').text
+    obj_struct['pose'] = obj.find('pose').text
+    obj_struct['truncated'] = int(obj.find('truncated').text)
+    obj_struct['difficult'] = int(obj.find('difficult').text)
+    bbox = obj.find('bndbox')
+    obj_struct['bbox'] = [int(bbox.find('xmin').text),
+                          int(bbox.find('ymin').text),
+                          int(bbox.find('xmax').text),
+                          int(bbox.find('ymax').text)]
+    objects.append(obj_struct)
+
+  return objects
+
+
+def voc_ap(rec, prec, use_07_metric=False):
+  """ ap = voc_ap(rec, prec, [use_07_metric])
+  Compute VOC AP given precision and recall.
+  If use_07_metric is true, uses the
+  VOC 07 11 point method (default:False).
+  """
+  if use_07_metric:
+    # 11 point metric
+    ap = 0.
+    for t in np.arange(0., 1.1, 0.1):
+      if np.sum(rec >= t) == 0:
+        p = 0
+      else:
+        p = np.max(prec[rec >= t])
+      ap = ap + p / 11.
+  else:
+    # correct AP calculation
+    # first append sentinel values at the end
+    mrec = np.concatenate(([0.], rec, [1.]))
+    mpre = np.concatenate(([0.], prec, [0.]))
+
+    # compute the precision envelope
+    for i in range(mpre.size - 1, 0, -1):
+      mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
+
+    # to calculate area under PR curve, look for points
+    # where X axis (recall) changes value
+    i = np.where(mrec[1:] != mrec[:-1])[0]
+
+    # and sum (\Delta recall) * prec
+    ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
+  return ap
+
+
+def voc_eval(detpath,
+             annopath,
+             imagesetfile,
+             classname,
+             cachedir,
+             ovthresh=0.5,
+             use_07_metric=False,
+             use_diff=False):
+  """rec, prec, ap = voc_eval(detpath,
+                              annopath,
+                              imagesetfile,
+                              classname,
+                              [ovthresh],
+                              [use_07_metric])
+
+  Top level function that does the PASCAL VOC evaluation.
+
+  detpath: Path to detections
+      detpath.format(classname) should produce the detection results file.
+  annopath: Path to annotations
+      annopath.format(imagename) should be the xml annotations file.
+  imagesetfile: Text file containing the list of images, one image per line.
+  classname: Category name (duh)
+  cachedir: Directory for caching the annotations
+  [ovthresh]: Overlap threshold (default = 0.5)
+  [use_07_metric]: Whether to use VOC07's 11 point AP computation
+      (default False)
+  """
+  # assumes detections are in detpath.format(classname)
+  # assumes annotations are in annopath.format(imagename)
+  # assumes imagesetfile is a text file with each line an image name
+  # cachedir caches the annotations in a pickle file
+
+  # first load gt
+  if not os.path.isdir(cachedir):
+    os.mkdir(cachedir)
+  cachefile = os.path.join(cachedir, '%s_annots.pkl' % imagesetfile)
+  # read list of images
+  with open(imagesetfile, 'r') as f:
+    lines = f.readlines()
+  imagenames = [x.strip() for x in lines]
+
+  if not os.path.isfile(cachefile):
+    # load annotations
+    recs = {}
+    for i, imagename in enumerate(imagenames):
+      recs[imagename] = parse_rec(annopath.format(imagename))
+      if i % 100 == 0:
+        print('Reading annotation for {:d}/{:d}'.format(
+          i + 1, len(imagenames)))
+    # save
+    print('Saving cached annotations to {:s}'.format(cachefile))
+    with open(cachefile, 'wb') as f:
+      pickle.dump(recs, f)
+  else:
+    # load
+    with open(cachefile, 'rb') as f:
+      try:
+        recs = pickle.load(f)
+      except:
+        recs = pickle.load(f, encoding='bytes')
+
+  # extract gt objects for this class
+  class_recs = {}
+  npos = 0
+  for imagename in imagenames:
+    R = [obj for obj in recs[imagename] if obj['name'] == classname]
+    bbox = np.array([x['bbox'] for x in R])
+    if use_diff:
+      difficult = np.array([False for x in R]).astype(np.bool)
+    else:
+      difficult = np.array([x['difficult'] for x in R]).astype(np.bool)
+    det = [False] * len(R)
+    npos = npos + sum(~difficult)
+    class_recs[imagename] = {'bbox': bbox,
+                             'difficult': difficult,
+                             'det': det}
+
+  # read dets
+  detfile = detpath.format(classname)
+  with open(detfile, 'r') as f:
+    lines = f.readlines()
+
+  splitlines = [x.strip().split(' ') for x in lines]
+  image_ids = [x[0] for x in splitlines]
+  confidence = np.array([float(x[1]) for x in splitlines])
+  BB = np.array([[float(z) for z in x[2:]] for x in splitlines])
+
+  nd = len(image_ids)
+  tp = np.zeros(nd)
+  fp = np.zeros(nd)
+
+  if BB.shape[0] > 0:
+    # sort by confidence
+    sorted_ind = np.argsort(-confidence)
+    sorted_scores = np.sort(-confidence)
+    BB = BB[sorted_ind, :]
+    image_ids = [image_ids[x] for x in sorted_ind]
+
+    # go down dets and mark TPs and FPs
+    for d in range(nd):
+      R = class_recs[image_ids[d]]
+      bb = BB[d, :].astype(float)
+      ovmax = -np.inf
+      BBGT = R['bbox'].astype(float)
+
+      if BBGT.size > 0:
+        # compute overlaps
+        # intersection
+        ixmin = np.maximum(BBGT[:, 0], bb[0])
+        iymin = np.maximum(BBGT[:, 1], bb[1])
+        ixmax = np.minimum(BBGT[:, 2], bb[2])
+        iymax = np.minimum(BBGT[:, 3], bb[3])
+        iw = np.maximum(ixmax - ixmin + 1., 0.)
+        ih = np.maximum(iymax - iymin + 1., 0.)
+        inters = iw * ih
+
+        # union
+        uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
+               (BBGT[:, 2] - BBGT[:, 0] + 1.) *
+               (BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)
+
+        overlaps = inters / uni
+        ovmax = np.max(overlaps)
+        jmax = np.argmax(overlaps)
+
+      if ovmax > ovthresh:
+        if not R['difficult'][jmax]:
+          if not R['det'][jmax]:
+            tp[d] = 1.
+            R['det'][jmax] = 1
+          else:
+            fp[d] = 1.
+      else:
+        fp[d] = 1.
+
+  # compute precision recall
+  fp = np.cumsum(fp)
+  tp = np.cumsum(tp)
+  rec = tp / float(npos)
+  # avoid divide by zero in case the first detection matches a difficult
+  # ground truth
+  prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
+  ap = voc_ap(rec, prec, use_07_metric)
+
+  return rec, prec, ap
diff --git a/lib/faster_rcnn.egg-info/PKG-INFO b/lib/faster_rcnn.egg-info/PKG-INFO
new file mode 100644
index 0000000..c59e971
--- /dev/null
+++ b/lib/faster_rcnn.egg-info/PKG-INFO
@@ -0,0 +1,10 @@
+Metadata-Version: 1.0
+Name: faster-rcnn
+Version: 0.1
+Summary: Weakly Supervised Obejct Detection
+Home-page: UNKNOWN
+Author: GaoWei
+Author-email: UNKNOWN
+License: UNKNOWN
+Description: UNKNOWN
+Platform: UNKNOWN
diff --git a/lib/faster_rcnn.egg-info/SOURCES.txt b/lib/faster_rcnn.egg-info/SOURCES.txt
new file mode 100644
index 0000000..fc53b63
--- /dev/null
+++ b/lib/faster_rcnn.egg-info/SOURCES.txt
@@ -0,0 +1,73 @@
+setup_cuda.py
+datasets/__init__.py
+datasets/coco.py
+datasets/ds_utils.py
+datasets/factory.py
+datasets/imdb.py
+datasets/pascal_voc.py
+datasets/voc_eval.py
+faster_rcnn.egg-info/PKG-INFO
+faster_rcnn.egg-info/SOURCES.txt
+faster_rcnn.egg-info/dependency_links.txt
+faster_rcnn.egg-info/not-zip-safe
+faster_rcnn.egg-info/top_level.txt
+layer_utils/__init__.py
+layer_utils/anchor_target_layer.py
+layer_utils/generate_anchors.py
+layer_utils/proposal_layer.py
+layer_utils/proposal_target_layer.py
+layer_utils/proposal_top_layer.py
+layer_utils/snippets.py
+layer_utils/roi_align/__init__.py
+layer_utils/roi_align/build.py
+layer_utils/roi_align/crop_and_resize.py
+layer_utils/roi_align/roi_align.py
+layer_utils/roi_align/_ext/__init__.py
+layer_utils/roi_align/_ext/crop_and_resize/__init__.py
+layer_utils/roi_pooling/__init__.py
+layer_utils/roi_pooling/build.py
+layer_utils/roi_pooling/_ext/__init__.py
+layer_utils/roi_pooling/_ext/roi_pooling/__init__.py
+layer_utils/roi_pooling/functions/__init__.py
+layer_utils/roi_pooling/functions/roi_pool.py
+layer_utils/roi_pooling/modules/__init__.py
+layer_utils/roi_pooling/modules/roi_pool.py
+model/__init__.py
+model/apmetric.py
+model/bbox_transform.py
+model/config.py
+model/nms_wrapper.py
+model/test.py
+model/train_val.py
+nets/__init__.py
+nets/mobilenet_v1.py
+nets/network.py
+nets/resnet_v1.py
+nets/vgg16.py
+nms/__init__.py
+nms/build.py
+nms/pth_nms.py
+nms/_ext/__init__.py
+nms/_ext/nms/__init__.py
+ops/nms/src/nms_cpu.cpp
+ops/nms/src/nms_cuda.cpp
+ops/nms/src/nms_kernel.cu
+ops/nms/src/soft_nms_cpu.cpp
+ops/roi_align/src/roi_align_cuda.cpp
+ops/roi_align/src/roi_align_kernel.cu
+ops/roi_crop/src/roi_crop_cpu.cpp
+ops/roi_crop/src/roi_crop_cuda.cpp
+ops/roi_crop/src/roi_crop_kernel.cu
+ops/roi_pool/src/roi_pool_cuda.cpp
+ops/roi_pool/src/roi_pool_kernel.cu
+ops/roi_ring_pool/src/roi_ring_pool_cuda.cpp
+ops/roi_ring_pool/src/roi_ring_pool_kernel.cu
+roi_data_layer/__init__.py
+roi_data_layer/layer.py
+roi_data_layer/minibatch.py
+roi_data_layer/roidb.py
+utils/__init__.py
+utils/bbox.py
+utils/blob.py
+utils/timer.py
+utils/visualization.py
\ No newline at end of file
diff --git a/lib/faster_rcnn.egg-info/dependency_links.txt b/lib/faster_rcnn.egg-info/dependency_links.txt
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/lib/faster_rcnn.egg-info/dependency_links.txt
@@ -0,0 +1 @@
+
diff --git a/lib/faster_rcnn.egg-info/not-zip-safe b/lib/faster_rcnn.egg-info/not-zip-safe
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/lib/faster_rcnn.egg-info/not-zip-safe
@@ -0,0 +1 @@
+
diff --git a/lib/faster_rcnn.egg-info/top_level.txt b/lib/faster_rcnn.egg-info/top_level.txt
new file mode 100644
index 0000000..1630b74
--- /dev/null
+++ b/lib/faster_rcnn.egg-info/top_level.txt
@@ -0,0 +1,8 @@
+datasets
+layer_utils
+model
+nets
+nms
+ops
+roi_data_layer
+utils
diff --git a/lib/make.sh b/lib/make.sh
new file mode 100644
index 0000000..1d3f18f
--- /dev/null
+++ b/lib/make.sh
@@ -0,0 +1,38 @@
+CUDA_ARCH="-gencode arch=compute_30,code=sm_30 \
+           -gencode arch=compute_35,code=sm_35 \
+           -gencode arch=compute_50,code=sm_50 \
+           -gencode arch=compute_52,code=sm_52 \
+           -gencode arch=compute_60,code=sm_60 \
+           -gencode arch=compute_61,code=sm_61 "
+
+# Build RoiPooling module
+
+cd layer_utils/roi_pooling/src
+echo "Compiling roi_pooling kernels by nvcc..."
+nvcc -c -o roi_pooling_kernel.cu.o roi_pooling_kernel.cu -x cu -Xcompiler -fPIC $CUDA_ARCH 
+cd ../
+python build.py
+cd ../../
+
+# Build RoiRingpooling module
+cd layer_utils/roi_ring_pooling/src
+echo "Compiling roi_ring_pooling kernels by nvcc"
+nvcc -c -o roi_ring_pooling_kernel.cu.o roi_ring_pooling_kernel.cu -x cu -Xcompiler -fPIC $CUDA_ARCH 
+cd ..
+python build.py
+
+# Build RoIAlign
+cd layer_utils/roi_align/src/cuda
+echo 'Compiling crop_and_resize kernels by nvcc...'
+nvcc -c -o crop_and_resize_kernel.cu.o crop_and_resize_kernel.cu -x cu -Xcompiler -fPIC $CUDA_ARCH
+cd ../../
+python build.py
+cd ../../
+
+# Build NMS
+cd nms/src/cuda
+echo "Compiling nms kernels by nvcc..."
+nvcc -c -o nms_kernel.cu.o nms_kernel.cu -x cu -Xcompiler -fPIC $CUDA_ARCH
+cd ../../
+python build.py
+cd ../
diff --git a/lib/make_cuda.sh b/lib/make_cuda.sh
new file mode 100644
index 0000000..23bfe54
--- /dev/null
+++ b/lib/make_cuda.sh
@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+export CUDA_PATH=/usr/local/cuda/
+#You may also want to ad the following
+export C_INCLUDE_PATH=/usr/local/cuda/include
+python setup_cuda.py develop
\ No newline at end of file
diff --git a/lib/model/__init__.py b/lib/model/__init__.py
new file mode 100644
index 0000000..d63bc18
--- /dev/null
+++ b/lib/model/__init__.py
@@ -0,0 +1 @@
+from . import config
diff --git a/lib/model/__pycache__/__init__.cpython-36.pyc b/lib/model/__pycache__/__init__.cpython-36.pyc
new file mode 100644
index 0000000..cdb52eb
Binary files /dev/null and b/lib/model/__pycache__/__init__.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/__init__.cpython-37.pyc b/lib/model/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..16984bd
Binary files /dev/null and b/lib/model/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/model/__pycache__/apmetric.cpython-36.pyc b/lib/model/__pycache__/apmetric.cpython-36.pyc
new file mode 100644
index 0000000..fde150f
Binary files /dev/null and b/lib/model/__pycache__/apmetric.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/bbox_transform.cpython-36.pyc b/lib/model/__pycache__/bbox_transform.cpython-36.pyc
new file mode 100644
index 0000000..821a773
Binary files /dev/null and b/lib/model/__pycache__/bbox_transform.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/bbox_transform.cpython-37.pyc b/lib/model/__pycache__/bbox_transform.cpython-37.pyc
new file mode 100644
index 0000000..a5dbbc0
Binary files /dev/null and b/lib/model/__pycache__/bbox_transform.cpython-37.pyc differ
diff --git a/lib/model/__pycache__/config.cpython-36.pyc b/lib/model/__pycache__/config.cpython-36.pyc
new file mode 100644
index 0000000..b648a7e
Binary files /dev/null and b/lib/model/__pycache__/config.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/config.cpython-37.pyc b/lib/model/__pycache__/config.cpython-37.pyc
new file mode 100644
index 0000000..fbbb4b3
Binary files /dev/null and b/lib/model/__pycache__/config.cpython-37.pyc differ
diff --git a/lib/model/__pycache__/nms_wrapper.cpython-36.pyc b/lib/model/__pycache__/nms_wrapper.cpython-36.pyc
new file mode 100644
index 0000000..93ed370
Binary files /dev/null and b/lib/model/__pycache__/nms_wrapper.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/nms_wrapper.cpython-37.pyc b/lib/model/__pycache__/nms_wrapper.cpython-37.pyc
new file mode 100644
index 0000000..2aa3847
Binary files /dev/null and b/lib/model/__pycache__/nms_wrapper.cpython-37.pyc differ
diff --git a/lib/model/__pycache__/test.cpython-36.pyc b/lib/model/__pycache__/test.cpython-36.pyc
new file mode 100644
index 0000000..7cf5e61
Binary files /dev/null and b/lib/model/__pycache__/test.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/test.cpython-37.pyc b/lib/model/__pycache__/test.cpython-37.pyc
new file mode 100644
index 0000000..3659ca5
Binary files /dev/null and b/lib/model/__pycache__/test.cpython-37.pyc differ
diff --git a/lib/model/__pycache__/train_val.cpython-36.pyc b/lib/model/__pycache__/train_val.cpython-36.pyc
new file mode 100644
index 0000000..ba97290
Binary files /dev/null and b/lib/model/__pycache__/train_val.cpython-36.pyc differ
diff --git a/lib/model/__pycache__/train_val.cpython-37.pyc b/lib/model/__pycache__/train_val.cpython-37.pyc
new file mode 100644
index 0000000..8db87af
Binary files /dev/null and b/lib/model/__pycache__/train_val.cpython-37.pyc differ
diff --git a/lib/model/apmetric.py b/lib/model/apmetric.py
new file mode 100644
index 0000000..fb88b7e
--- /dev/null
+++ b/lib/model/apmetric.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Thu Nov 22 16:14:59 2018
+
+@author: vasgaoweithu
+"""
+
+import torch
+import numpy as np
+from copy import deepcopy
+import math
+
+class AveragePrecisionMeter(object):
+    def __init__(self, difficult_examples=False):
+        super(AveragePrecisionMeter, self).__init__()
+        self.reset()
+        self.difficult_examples = difficult_examples
+
+    def reset(self):
+        self.scores = torch.FloatTensor(torch.FloatStorage())
+        self.targets = torch.LongTensor(torch.LongStorage())
+
+    def add(self, output, target):
+        if not torch.is_tensor(output):
+            output = torch.from_numpy(output)
+        if not torch.is_tensor(target):
+            target = torch.from_numpy(target)
+
+        if output.dim() == 1:
+            output = output.view(-1, 1)
+        else:
+            assert output.dim() == 2, \
+                'wrong output size (should be 1D or 2D with one column \
+                per class)'
+        if target.dim() == 1:
+            target = target.view(-1, 1)
+        else:
+            assert target.dim() == 2, \
+                'wrong target size (should be 1D or 2D with one column \
+                per class)'
+        if self.scores.numel() > 0:
+            assert target.size(1) == self.targets.size(1), \
+                'dimensions for output should match previously added examples.'
+
+        # make sure storage is of sufficient size
+        if self.scores.storage().size() < self.scores.numel() + output.numel():
+            new_size = math.ceil(self.scores.storage().size() * 1.5)
+            self.scores.storage().resize_(int(new_size + output.numel()))
+            self.targets.storage().resize_(int(new_size + output.numel()))
+
+        # store scores and targets
+        offset = self.scores.size(0) if self.scores.dim() > 0 else 0
+        self.scores.resize_(offset + output.size(0), output.size(1))
+        self.targets.resize_(offset + target.size(0), target.size(1))
+        self.scores.narrow(0, offset, output.size(0)).copy_(output)
+        self.targets.narrow(0, offset, target.size(0)).copy_(target)
+
+    def value(self):
+        if self.scores.numel() == 0:
+            return 0
+        ap = torch.zeros(self.scores.size(1))
+        rg = torch.arange(1, self.scores.size(0)).float()
+
+        # compute average precision for each class
+        for k in range(self.scores.size(1)):
+            # sort scores
+            scores = self.scores[:, k]
+            targets = self.targets[:, k]
+            
+            # compute average precision
+            ap[k] = AveragePrecisionMeter.average_precision(scores, targets, self.difficult_examples)
+        return ap
+
+    @staticmethod
+    def average_precision(output, target, difficult_examples=True):
+
+        # sort examples
+        sorted, indices = torch.sort(output, dim=0, descending=True)
+
+        # Computes prec@i
+        pos_count = 0.
+        total_count = 0.
+        precision_at_i = 0.
+        for i in indices:
+            label = target[i]
+            if difficult_examples and label == 0:
+                continue
+            if label == 1:
+                pos_count += 1
+            total_count += 1
+            if label == 1:
+                precision_at_i += pos_count / total_count
+        precision_at_i /= pos_count
+        return precision_at_i
\ No newline at end of file
diff --git a/lib/model/bbox_transform.py b/lib/model/bbox_transform.py
new file mode 100644
index 0000000..66916a8
--- /dev/null
+++ b/lib/model/bbox_transform.py
@@ -0,0 +1,80 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import torch
+
+def bbox_transform(ex_rois, gt_rois):
+  ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
+  ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
+  ex_ctr_x = ex_rois[:, 0] + 0.5 * ex_widths
+  ex_ctr_y = ex_rois[:, 1] + 0.5 * ex_heights
+
+  gt_widths = gt_rois[:, 2] - gt_rois[:, 0] + 1.0
+  gt_heights = gt_rois[:, 3] - gt_rois[:, 1] + 1.0
+  gt_ctr_x = gt_rois[:, 0] + 0.5 * gt_widths
+  gt_ctr_y = gt_rois[:, 1] + 0.5 * gt_heights
+
+  targets_dx = (gt_ctr_x - ex_ctr_x) / ex_widths
+  targets_dy = (gt_ctr_y - ex_ctr_y) / ex_heights
+  targets_dw = torch.log(gt_widths / ex_widths)
+  targets_dh = torch.log(gt_heights / ex_heights)
+
+  targets = torch.stack(
+    (targets_dx, targets_dy, targets_dw, targets_dh), 1)
+  return targets
+
+
+def bbox_transform_inv(boxes, deltas):
+  # Input should be both tensor or both Variable and on the same device
+  if len(boxes) == 0:
+    return deltas.detach() * 0
+
+  widths = boxes[:, 2] - boxes[:, 0] + 1.0
+  heights = boxes[:, 3] - boxes[:, 1] + 1.0
+  ctr_x = boxes[:, 0] + 0.5 * widths
+  ctr_y = boxes[:, 1] + 0.5 * heights
+
+  dx = deltas[:, 0::4]
+  dy = deltas[:, 1::4]
+  dw = deltas[:, 2::4]
+  dh = deltas[:, 3::4]
+  
+  pred_ctr_x = dx * widths.unsqueeze(1) + ctr_x.unsqueeze(1)
+  pred_ctr_y = dy * heights.unsqueeze(1) + ctr_y.unsqueeze(1)
+  pred_w = torch.exp(dw) * widths.unsqueeze(1)
+  pred_h = torch.exp(dh) * heights.unsqueeze(1)
+
+  pred_boxes = torch.cat(\
+    [_.unsqueeze(2) for _ in [pred_ctr_x - 0.5 * pred_w,\
+                              pred_ctr_y - 0.5 * pred_h,\
+                              pred_ctr_x + 0.5 * pred_w,\
+                              pred_ctr_y + 0.5 * pred_h]], 2).view(len(boxes), -1)
+
+  return pred_boxes
+
+
+def clip_boxes(boxes, im_shape):
+  """
+  Clip boxes to image boundaries.
+  boxes must be tensor or Variable, im_shape can be anything but Variable
+  """
+
+  if not hasattr(boxes, 'data'):
+    boxes_ = boxes.numpy()
+
+  boxes = boxes.view(boxes.size(0), -1, 4)
+  boxes = torch.stack(\
+    [boxes[:,:,0].clamp(0, im_shape[1] - 1),
+     boxes[:,:,1].clamp(0, im_shape[0] - 1),
+     boxes[:,:,2].clamp(0, im_shape[1] - 1),
+     boxes[:,:,3].clamp(0, im_shape[0] - 1)], 2).view(boxes.size(0), -1)
+
+  return boxes
diff --git a/lib/model/config.py b/lib/model/config.py
new file mode 100644
index 0000000..b8febb6
--- /dev/null
+++ b/lib/model/config.py
@@ -0,0 +1,394 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import os.path as osp
+import numpy as np
+# `pip install easydict` if you don't have it
+from easydict import EasyDict as edict
+
+__C = edict()
+# Consumers can get config by:
+#   from fast_rcnn_config import cfg
+cfg = __C
+
+#
+# Training options
+#
+__C.TRAIN = edict()
+
+# Initial learning rate
+__C.TRAIN.LEARNING_RATE = 0.001
+
+# Momentum
+__C.TRAIN.MOMENTUM = 0.9
+
+# Weight decay, for regularization
+__C.TRAIN.WEIGHT_DECAY = 0.0001
+
+# Factor for reducing the learning rate
+__C.TRAIN.GAMMA = 0.1
+
+# Step size for reducing the learning rate, currently only support one step
+__C.TRAIN.STEPSIZE = [60000]
+
+# Iteration intervals for showing the loss during training, on command line interface
+__C.TRAIN.DISPLAY = 10
+
+# Whether to double the learning rate for bias
+__C.TRAIN.DOUBLE_BIAS = True
+
+# Whether to initialize the weights with truncated normal distribution 
+__C.TRAIN.TRUNCATED = False
+
+# Whether to have weight decay on bias as well
+__C.TRAIN.BIAS_DECAY = False
+
+# Whether to add ground truth boxes to the pool when sampling regions
+__C.TRAIN.USE_GT = False
+
+# Whether to use aspect-ratio grouping of training images, introduced merely for saving
+# GPU memory
+__C.TRAIN.ASPECT_GROUPING = False
+
+# The number of snapshots kept, older ones are deleted to save space
+__C.TRAIN.SNAPSHOT_KEPT = 3
+
+# The time interval for saving tensorflow summaries
+__C.TRAIN.SUMMARY_INTERVAL = 180
+
+# Scale to use during training (can list multiple scales)
+# The scale is the pixel size of an image's shortest side
+__C.TRAIN.SCALES = (600,)
+
+# Max pixel size of the longest side of a scaled input image
+__C.TRAIN.MAX_SIZE = 1000
+
+# Images to use per minibatch
+__C.TRAIN.IMS_PER_BATCH = 1
+
+# Minibatch size (number of regions of interest [ROIs])
+__C.TRAIN.BATCH_SIZE = 128
+
+# Fraction of minibatch that is labeled foreground (i.e. class > 0)
+__C.TRAIN.FG_FRACTION = 0.25
+
+# Overlap threshold for a ROI to be considered foreground (if >= FG_THRESH)
+__C.TRAIN.FG_THRESH = 0.55
+
+# Overlap threshold for a ROI to be considered background (class = 0 if
+# overlap in [LO, HI))
+__C.TRAIN.BG_THRESH_HI = 0.45
+__C.TRAIN.BG_THRESH_LO = 0.1
+
+__C.TRAIN.MIL_FG_THRESH = 0.6
+__C.TRAIN.MIL_BG_THRESH_LO = 0.1
+__C.TRAIN.MIL_BG_THRESH_HI = 0.4
+__C.TRAIN.MIL_RECURRENT_STEP = 30000
+__C.TRAIN.MIL_RECURRECT_WEIGHT = 0.5
+
+__C.TRAIN.MIL_NUM_FG = 32
+__C.TRAIN.MIL_NUM_BG = 96
+__C.TRAIN.MIL_BATCHSIZE = 128
+
+
+# Use horizontally-flipped images during training?
+__C.TRAIN.USE_FLIPPED = True
+
+# Train bounding-box regressors
+__C.TRAIN.BBOX_REG = True
+
+# Overlap required between a ROI and ground-truth box in order for that ROI to
+# be used as a bounding-box regression training example
+__C.TRAIN.BBOX_THRESH = 0.5
+
+# Iterations between snapshots
+__C.TRAIN.SNAPSHOT_ITERS = 5000
+__C.TRAIN.STEP_ITERS = 50000  # voc07
+
+# solver.prototxt specifies the snapshot path prefix, this adds an optional
+# infix to yield the path: <prefix>[_<infix>]_iters_XYZ.caffemodel
+__C.TRAIN.SNAPSHOT_PREFIX = 'res101_faster_rcnn'
+
+# Normalize the targets (subtract empirical mean, divide by empirical stddev)
+__C.TRAIN.BBOX_NORMALIZE_TARGETS = True
+
+# Deprecated (inside weights)
+__C.TRAIN.BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+
+# Normalize the targets using "precomputed" (or made up) means and stdevs
+# (BBOX_NORMALIZE_TARGETS must also be True)
+__C.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED = True
+
+__C.TRAIN.BBOX_NORMALIZE_MEANS = (0.0, 0.0, 0.0, 0.0)
+
+__C.TRAIN.BBOX_NORMALIZE_STDS = (0.1, 0.1, 0.2, 0.2)
+
+# Train using these proposals
+__C.TRAIN.PROPOSAL_METHOD = 'gt'
+
+# Make minibatches from images that have similar aspect ratios (i.e. both
+# tall and thin or both short and wide) in order to avoid wasting computation
+# on zero-padding.
+
+# Use RPN to detect objects
+__C.TRAIN.HAS_RPN = True
+
+# IOU >= thresh: positive example
+__C.TRAIN.RPN_POSITIVE_OVERLAP = 0.7
+
+# IOU < thresh: negative example
+__C.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3
+
+# If an anchor satisfied by positive and negative conditions set to negative
+__C.TRAIN.RPN_CLOBBER_POSITIVES = False
+
+# Max number of foreground examples
+__C.TRAIN.RPN_FG_FRACTION = 0.5
+
+# Total number of examples
+__C.TRAIN.RPN_BATCHSIZE = 256
+
+# NMS threshold used on RPN proposals
+__C.TRAIN.RPN_NMS_THRESH = 0.7
+
+# Number of top scoring boxes to keep before apply NMS to RPN proposals
+__C.TRAIN.RPN_PRE_NMS_TOP_N = 12000
+
+# Number of top scoring boxes to keep after applying NMS to RPN proposals
+__C.TRAIN.RPN_POST_NMS_TOP_N = 2000
+
+# Deprecated (outside weights)
+__C.TRAIN.RPN_BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+
+# Give the positive RPN examples weight of p * 1 / {num positives}
+# and give negatives a weight of (1 - p)
+# Set to -1.0 to use uniform example weighting
+__C.TRAIN.RPN_POSITIVE_WEIGHT = -1.0
+
+# Whether to use all ground truth bounding boxes for training, 
+# For COCO, setting USE_ALL_GT to False will exclude boxes that are flagged as ''iscrowd''
+__C.TRAIN.USE_ALL_GT = True
+
+#
+# Testing options
+#
+__C.TEST = edict()
+
+# Scale to use during testing (can NOT list multiple scales)
+# The scale is the pixel size of an image's shortest side
+__C.TEST.SCALES = (600,)
+
+# Max pixel size of the longest side of a scaled input image
+__C.TEST.MAX_SIZE = 1000
+
+# Overlap threshold used for non-maximum suppression (suppress boxes with
+# IoU >= this threshold)
+__C.TEST.NMS = 0.3
+
+# Experimental: treat the (K+1) units in the cls_score layer as linear
+# predictors (trained, eg, with one-vs-rest SVMs).
+__C.TEST.SVM = False
+
+# Test using bounding-box regressors
+__C.TEST.BBOX_REG = True
+
+# Propose boxes
+__C.TEST.HAS_RPN = False
+
+# Test using these proposals
+__C.TEST.PROPOSAL_METHOD = 'gt'
+
+## NMS threshold used on RPN proposals
+__C.TEST.RPN_NMS_THRESH = 0.7
+
+# Number of top scoring boxes to keep before apply NMS to RPN proposals
+__C.TEST.RPN_PRE_NMS_TOP_N = 6000
+
+# Number of top scoring boxes to keep after applying NMS to RPN proposals
+__C.TEST.RPN_POST_NMS_TOP_N = 300
+
+# Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale)
+# __C.TEST.RPN_MIN_SIZE = 16
+
+# Testing mode, default to be 'nms', 'top' is slower but better
+# See report for details
+__C.TEST.MODE = 'nms'
+
+# Only useful when TEST.MODE is 'top', specifies the number of top proposals to select
+__C.TEST.RPN_TOP_N = 5000
+
+#
+# ResNet options
+#
+
+__C.RESNET = edict()
+
+# Option to set if max-pooling is appended after crop_and_resize. 
+# if true, the region will be resized to a square of 2xPOOLING_SIZE, 
+# then 2x2 max-pooling is applied; otherwise the region will be directly
+# resized to a square of POOLING_SIZE
+__C.RESNET.MAX_POOL = False
+
+# Number of fixed blocks during training, by default the first of all 4 blocks is fixed
+# Range: 0 (none) to 3 (all)
+__C.RESNET.FIXED_BLOCKS = 1
+
+#
+# MobileNet options
+#
+
+__C.MOBILENET = edict()
+
+# Whether to regularize the depth-wise filters during training
+__C.MOBILENET.REGU_DEPTH = False
+
+# Number of fixed layers during training, by default the bottom 5 of 14 layers is fixed
+# Range: 0 (none) to 12 (all)
+__C.MOBILENET.FIXED_LAYERS = 5
+
+# Weight decay for the mobilenet weights
+__C.MOBILENET.WEIGHT_DECAY = 0.00004
+
+# Depth multiplier
+__C.MOBILENET.DEPTH_MULTIPLIER = 1.
+
+#
+# MISC
+#
+
+# Pixel mean values (BGR order) as a (1, 1, 3) array
+# We use the same pixel mean for all networks even though it's not exactly what
+# they were trained with
+__C.PIXEL_MEANS = np.array([[[102.9801, 115.9465, 122.7717]]])
+
+# For reproducibility
+__C.RNG_SEED = 3
+
+# Root directory of project
+__C.ROOT_DIR = osp.abspath(osp.join(osp.dirname(__file__), '..', '..'))
+
+# Data directory
+__C.DATA_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'data'))
+
+# Name (or path to) the matlab executable
+__C.MATLAB = 'matlab'
+
+# Place outputs under an experiments directory
+__C.EXP_DIR = 'default'
+
+# Use GPU implementation of non-maximum suppression
+__C.USE_GPU_NMS = True
+
+# Default pooling mode
+__C.POOLING_MODE = 'crop'
+
+# Size of the pooled region after RoI pooling
+__C.POOLING_SIZE = 7
+
+# Anchor scales for RPN
+__C.ANCHOR_SCALES = [8,16,32]
+
+# Anchor ratios for RPN
+__C.ANCHOR_RATIOS = [0.5,1,2]
+
+# Number of filters for the RPN layer
+__C.RPN_CHANNELS = 512
+
+
+def get_output_dir(imdb, weights_filename):
+  """Return the directory where experimental artifacts are placed.
+  If the directory does not exist, it is created.
+
+  A canonical path is built using the name from an imdb and a network
+  (if not None).
+  """
+  outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'output', __C.EXP_DIR, imdb.name))
+  if weights_filename is None:
+    weights_filename = 'default'
+  outdir = osp.join(outdir, weights_filename)
+  if not os.path.exists(outdir):
+    os.makedirs(outdir)
+  return outdir
+
+
+def get_output_tb_dir(imdb, weights_filename):
+  """Return the directory where tensorflow summaries are placed.
+  If the directory does not exist, it is created.
+
+  A canonical path is built using the name from an imdb and a network
+  (if not None).
+  """
+  outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'tensorboard', __C.EXP_DIR, imdb.name))
+  if weights_filename is None:
+    weights_filename = 'default'
+  outdir = osp.join(outdir, weights_filename)
+  if not os.path.exists(outdir):
+    os.makedirs(outdir)
+  return outdir
+
+
+def _merge_a_into_b(a, b):
+  """Merge config dictionary a into config dictionary b, clobbering the
+  options in b whenever they are also specified in a.
+  """
+  if type(a) is not edict:
+    return
+
+  for k, v in a.items():
+    # a must specify keys that are in b
+    if k not in b:
+      raise KeyError('{} is not a valid config key'.format(k))
+
+    # the types must match, too
+    old_type = type(b[k])
+    if old_type is not type(v):
+      if isinstance(b[k], np.ndarray):
+        v = np.array(v, dtype=b[k].dtype)
+      else:
+        raise ValueError(('Type mismatch ({} vs. {}) '
+                          'for config key: {}').format(type(b[k]),
+                                                       type(v), k))
+
+    # recursively merge dicts
+    if type(v) is edict:
+      try:
+        _merge_a_into_b(a[k], b[k])
+      except:
+        print(('Error under config key: {}'.format(k)))
+        raise
+    else:
+      b[k] = v
+
+
+def cfg_from_file(filename):
+  """Load a config file and merge it into the default options."""
+  import yaml
+  with open(filename, 'r') as f:
+    yaml_cfg = edict(yaml.load(f))
+
+  _merge_a_into_b(yaml_cfg, __C)
+
+
+def cfg_from_list(cfg_list):
+  """Set config keys via list (e.g., from command line)."""
+  from ast import literal_eval
+  assert len(cfg_list) % 2 == 0
+  for k, v in zip(cfg_list[0::2], cfg_list[1::2]):
+    key_list = k.split('.')
+    d = __C
+    for subkey in key_list[:-1]:
+      assert subkey in d
+      d = d[subkey]
+    subkey = key_list[-1]
+    assert subkey in d
+    try:
+      value = literal_eval(v)
+    except:
+      # handle the case when v is a string literal
+      value = v
+    assert type(value) == type(d[subkey]), \
+      'type {} does not match original type {}'.format(
+        type(value), type(d[subkey]))
+    d[subkey] = value
diff --git a/lib/model/nms_wrapper.py b/lib/model/nms_wrapper.py
new file mode 100644
index 0000000..3e45e6a
--- /dev/null
+++ b/lib/model/nms_wrapper.py
@@ -0,0 +1,17 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from nms.pth_nms import pth_nms
+
+
+def nms(dets, thresh):
+  """Dispatch to either CPU or GPU NMS implementations.
+  Accept dets as tensor"""
+  return pth_nms(dets, thresh)
diff --git a/lib/model/test.py b/lib/model/test.py
new file mode 100644
index 0000000..600085d
--- /dev/null
+++ b/lib/model/test.py
@@ -0,0 +1,295 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import cv2
+import numpy as np
+
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+import os
+import math
+
+from utils.timer import Timer
+from ops.nms import nms
+from utils.blob import im_list_to_blob
+
+from model.config import cfg, get_output_dir
+from model.bbox_transform import clip_boxes, bbox_transform_inv
+
+import torch
+
+from model.apmetric import AveragePrecisionMeter
+
+
+def _get_image_blob(im):
+    """Converts an image into a network input.
+    Arguments:
+      im (ndarray): a color image in BGR order
+    Returns:
+      blob (ndarray): a data blob holding an image pyramid
+      im_scale_factors (list): list of image scales (relative to im) used
+        in the image pyramid
+    """
+    im_orig = im.astype(np.float32, copy=True)
+    im_orig -= cfg.PIXEL_MEANS
+
+    im_shape = im_orig.shape
+    im_size_min = np.min(im_shape[0:2])
+    im_size_max = np.max(im_shape[0:2])
+
+    processed_ims = []
+    im_scale_factors = []
+    im_shape = []
+
+    for target_size in cfg.TEST.SCALES:
+        im_scale = float(target_size) / float(im_size_min)
+        # Prevent the biggest axis from being more than MAX_SIZE
+        if np.round(im_scale * im_size_max) > cfg.TEST.MAX_SIZE:
+            im_scale = float(cfg.TEST.MAX_SIZE) / float(im_size_max)
+        im = cv2.resize(im_orig, None, None, fx=im_scale, fy=im_scale,
+                        interpolation=cv2.INTER_LINEAR)
+        im_scale_factors.append(im_scale)
+        processed_ims.append(im)
+        im_shape.append(im.shape)
+
+    # Create a blob to hold the input images
+    blob = im_list_to_blob(processed_ims)
+
+    return blob, np.array(im_scale_factors), im_shape
+
+
+def _get_blobs(im):
+    """Convert an image and RoIs within that image into network inputs."""
+    blobs = {}
+    blobs['data'], im_scale_factors, im_shape = _get_image_blob(im)
+
+    return blobs, im_scale_factors, im_shape
+
+
+def _clip_boxes(boxes, im_shape):
+    """Clip boxes to image boundaries."""
+    # x1 >= 0
+    boxes[:, 0::4] = np.maximum(boxes[:, 0::4], 0)
+    # y1 >= 0
+    boxes[:, 1::4] = np.maximum(boxes[:, 1::4], 0)
+    # x2 < im_shape[1]
+    boxes[:, 2::4] = np.minimum(boxes[:, 2::4], im_shape[1] - 1)
+    # y2 < im_shape[0]
+    boxes[:, 3::4] = np.minimum(boxes[:, 3::4], im_shape[0] - 1)
+    return boxes
+
+def _shuffle_boxes(boxes, w, h):
+    num_ss = boxes.shape[0]
+    jet_scale = np.random.uniform(0.95, 1.15, num_ss)
+    jet_center_w = np.random.uniform(-0.05, 0.05, num_ss)
+    jet_center_h = np.random.uniform(-0.05, 0.05, num_ss)
+    ss_box_jet = boxes[:, 1:].copy()
+    widths = ss_box_jet[:, 2] - ss_box_jet[:, 0] + 1.0
+    heights = ss_box_jet[:, 3] - ss_box_jet[:, 1] + 1.0
+    ctr_x = ss_box_jet[:, 0] + 0.5 * widths + jet_center_w * widths
+    ctr_y = ss_box_jet[:, 1] + 0.5 * heights + jet_center_h * heights
+    widths_new = widths * jet_scale
+    heights_new = heights * jet_scale
+    boxes[:, 1] = ctr_x - widths_new * 0.5
+    boxes[:, 2] = ctr_y - heights_new * 0.5
+    boxes[:, 3] = ctr_x + widths_new * 0.5
+    boxes[:, 4] = ctr_y + heights_new * 0.5
+    boxes[:, 1] = np.clip(boxes[:, 1], 0, w - 1)
+    boxes[:, 2] = np.clip(boxes[:, 2], 0, h - 1)
+    boxes[:, 3] = np.clip(boxes[:, 3], 0, w - 1)
+    boxes[:, 4] = np.clip(boxes[:, 4], 0, h - 1)
+    return boxes
+
+def _rescale_boxes(boxes, inds, scales):
+    """Rescale boxes according to image rescaling."""
+    for i in range(boxes.shape[0]):
+        boxes[i, :] = boxes[i, :] / scales[int(inds[i])]
+
+    return boxes
+
+def get_ss_boxes(roidb_i, im_scales):
+    ss_inds = np.where(roidb_i['gt_classes'] == -1)[0]  # remove gt_rois in ss_boxes
+    ss_boxes = np.empty((len(ss_inds), 5), dtype=np.float32)
+    ss_boxes[:, 1:] = roidb_i['boxes'][ss_inds, :] * im_scales
+    ss_boxes[:, 0] = 0
+    return ss_boxes
+
+
+def get_flipper_boxes(ss_box, width):
+    oldx1 = ss_box[:, 1].copy()
+    oldx2 = ss_box[:, 3].copy()
+    ss_box[:, 1] = width - oldx2 - 1
+    ss_box[:, 3] = width - oldx1 - 1
+    return ss_box
+
+
+def im_detect(net, im, roidb_i):
+    blobs, im_scales, im_shape = _get_blobs(im)
+    # ss_boxes = get_ss_boxes(roidb_i, im_scales)
+    # assert len(im_scales) == 1, "Only single-image batch implemented"
+    # assert len(im_scales) == 5, "multiple scale test"
+
+    im_blob_scales = blobs['data']
+
+    score_list = []
+    pred_box_list = []
+    det_cls_list = []
+
+    for index, scale in enumerate(im_scales):
+        im_blob = im_blob_scales[index]
+        ss_boxes = get_ss_boxes(roidb_i, scale)
+        h, w = im_shape[index][0], im_shape[index][1]
+        ss_boxes_2 = ss_boxes.copy()
+        ss_boxes = _shuffle_boxes(ss_boxes, w, h)
+        ss_boxes_2 = _shuffle_boxes(ss_boxes_2, w, h)
+
+        im_blob_flip = im_blob.copy()[:, ::-1, :]
+        ss_boxes_flip = get_flipper_boxes(ss_boxes_2.copy(), im_blob.shape[1])
+        for k in range(2):   # input and its flip
+            if k % 2 == 1:
+                im_blob_flip = im_blob_flip[np.newaxis, :, :, :]
+                img_info = np.array([im_blob.shape[1], im_blob.shape[2], scale], dtype=np.float32)
+                bbox_pred, rois, det_cls_prob, det_cls_prob_product, refine_prob_1, refine_prob_2 = net.test_image(im_blob_flip, img_info, ss_boxes_flip)
+                boxes = ss_boxes_2[:, 1:5] / scale
+            else:
+                im_blob = im_blob[np.newaxis, :, :, :]
+                img_info = np.array([im_blob.shape[1], im_blob.shape[2], scale], dtype=np.float32)
+                bbox_pred, rois, det_cls_prob, det_cls_prob_product, refine_prob_1, refine_prob_2 = net.test_image(im_blob, img_info, ss_boxes)
+                boxes = ss_boxes[:, 1:5] / scale
+
+            # scores = np.reshape((refine_prob_1 + refine_prob_2) / 2, [det_cls_prob_product.shape[0], -1])
+            scores = np.reshape(refine_prob_1, [det_cls_prob_product.shape[0], -1])
+            bbox_pred = np.reshape(bbox_pred, [bbox_pred.shape[0], -1])
+
+            if cfg.TEST.BBOX_REG:
+                box_deltas = bbox_pred
+                pred_boxes = bbox_transform_inv(torch.from_numpy(boxes), torch.from_numpy(box_deltas)).numpy()
+                pred_boxes = _clip_boxes(pred_boxes, im.shape)
+            else:
+                pred_boxes = np.tile(boxes, (1, scores.shape[1]))
+            score_list.append(scores)
+            pred_box_list.append(pred_boxes)
+            det_cls_list.append(det_cls_prob)
+
+    scores = np.array(score_list).mean(axis=0)
+    pred_boxes = np.array(pred_box_list).mean(axis=0)
+    det_cls_prob = np.array(det_cls_list).mean(axis=0)
+    target = np.reshape(roidb_i['image_level_labels'], (-1))
+    return scores, pred_boxes, det_cls_prob, target
+
+
+def apply_nms(all_boxes, thresh):
+    """Apply non-maximum suppression to all predicted boxes output by the
+    test_net method.
+    """
+    num_classes = len(all_boxes)
+    num_images = len(all_boxes[0])
+    nms_boxes = [[[] for _ in range(num_images)] for _ in range(num_classes)]
+    for cls_ind in range(num_classes):
+        for im_ind in range(num_images):
+            dets = all_boxes[cls_ind][im_ind]
+            if dets == []:
+                continue
+
+            x1 = dets[:, 0]
+            y1 = dets[:, 1]
+            x2 = dets[:, 2]
+            y2 = dets[:, 3]
+            scores = dets[:, 4]
+            inds = np.where((x2 > x1) & (y2 > y1))[0]
+            dets = dets[inds, :]
+            if dets == []:
+                continue
+
+            keep = nms(torch.from_numpy(dets), thresh).numpy()
+            if len(keep) == 0:
+                continue
+            nms_boxes[cls_ind][im_ind] = dets[keep, :].copy()
+    return nms_boxes
+
+
+def test_net(net, imdb, roidb, weights_filename, max_per_image=100, thresh=0.):
+    np.random.seed(cfg.RNG_SEED)
+    """Test a Fast R-CNN network on an image database."""
+    num_images = len(imdb.image_index)
+    # all detections are collected into:
+    #  all_boxes[cls][image] = N x 5 array of detections in
+    #  (x1, y1, x2, y2, score)
+    all_boxes = [[[] for _ in range(num_images)]
+                 for _ in range(imdb.num_classes)]
+
+    output_dir = get_output_dir(imdb, weights_filename)
+
+    # -------------------------------------------------------
+    ap_meter = AveragePrecisionMeter(difficult_examples=True)
+    ap_meter.reset()
+    # timers
+    _t = {'im_detect': Timer(), 'misc': Timer()}
+
+    for i in range(num_images):
+        im = cv2.imread(imdb.image_path_at(i))
+
+        _t['im_detect'].tic()
+        scores, boxes, det_cls_prob, target = im_detect(net, im, roidb[i])
+        _t['im_detect'].toc()
+
+        _t['misc'].tic()
+
+        output = np.reshape(det_cls_prob[:], (1, -1))
+        target = np.reshape(target[:], (1, -1))
+        ap_meter.add(output, target)
+
+        # skip j = 0, because it's the background class
+        for j in range(0, imdb.num_classes):
+            inds = np.where(scores[:, j] > thresh)[0]
+            cls_scores = scores[inds, j]
+            cls_boxes = boxes[inds, j * 4:(j + 1) * 4]
+            cls_dets = np.hstack((cls_boxes, cls_scores[:, np.newaxis])) \
+                .astype(np.float32, copy=False)
+            # keep = nms(torch.from_numpy(cls_dets), cfg.TEST.NMS).numpy() if cls_dets.size > 0 else []
+            # cls_dets = cls_dets[keep, :]
+            # all_boxes[j][i] = cls_dets
+            keep = nms(torch.from_numpy(cls_dets), cfg.TEST.NMS) if cls_dets.size > 0 else []
+            all_boxes[j][i] = keep[0].numpy()
+
+        # Limit to max_per_image detections *over all classes*
+        if max_per_image > 0:
+            image_scores = np.hstack([all_boxes[j][i][:, -1]
+                                      for j in range(1, imdb.num_classes)])
+            if len(image_scores) > max_per_image:
+                image_thresh = np.sort(image_scores)[-max_per_image]
+                for j in range(1, imdb.num_classes):
+                    keep = np.where(all_boxes[j][i][:, -1] >= image_thresh)[0]
+                    all_boxes[j][i] = all_boxes[j][i][keep, :]
+        _t['misc'].toc()
+
+        if i % 100 == 0:
+            print('im_detect: {:d}/{:d} {:.3f}s {:.3f}s'.format(i + 1, num_images, _t['im_detect'].average_time(),
+                                                                _t['misc'].average_time()))
+
+    ap = ap_meter.value().numpy()
+    print('the classification AP is ')
+    for index, cls in enumerate(imdb._classes):
+        if cls == '__background__':
+            continue
+        print(('AP for {} = {:.4f}'.format(cls, ap[index])))
+    print('__________________')
+    map = 100 * ap.mean()
+    print('the mAP is {:.4f}'.format(map))
+
+    det_file = os.path.join(output_dir, 'detections.pkl')
+    with open(det_file, 'wb') as f:
+        pickle.dump(all_boxes, f, pickle.HIGHEST_PROTOCOL)
+
+    print('Evaluating detections')
+    imdb.evaluate_detections(all_boxes, output_dir)
+
+
diff --git a/lib/model/train_val.py b/lib/model/train_val.py
new file mode 100644
index 0000000..e547fda
--- /dev/null
+++ b/lib/model/train_val.py
@@ -0,0 +1,378 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen and Zheqi He
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorboardX as tb
+
+from model.config import cfg
+import roi_data_layer.roidb as rdl_roidb
+from roi_data_layer.layer import RoIDataLayer
+import utils.timer
+try:
+  import cPickle as pickle
+except ImportError:
+  import pickle
+
+import torch
+import torch.optim as optim
+import numpy as np
+import os
+import glob
+import time
+import torch.nn.functional as F
+from torch import nn
+
+def update_learning_rate(optimizer, cur_lr, new_lr):
+  """Update learning rate"""
+  if cur_lr != new_lr:
+    ratio = _get_lr_change_ratio(cur_lr, new_lr)
+    param_keys = []
+    for ind, param_group in enumerate(optimizer.param_groups):
+      param_group['lr'] = new_lr
+      param_keys += param_group['params']
+
+def _get_lr_change_ratio(cur_lr, new_lr):
+    eps = 1e-10
+    ratio = np.max(
+        (new_lr / np.max((cur_lr, eps)), cur_lr / np.max((new_lr, eps)))
+    )
+    return ratio
+
+def scale_lr(optimizer, scale):
+  """Scale the learning rate of the optimizer"""
+  for param_group in optimizer.param_groups:
+    param_group['lr'] *= scale
+
+class SolverWrapper(object):
+  """
+    A wrapper class for the training process
+  """
+
+  def __init__(self, network, imdb, roidb, valroidb, output_dir, tbdir, pretrained_model=None):
+    self.net = network
+    self.imdb = imdb
+    self.roidb = roidb
+    self.valroidb = valroidb
+    self.output_dir = output_dir
+    self.tbdir = tbdir
+    # Simply put '_val' at the end to save the summaries from the validation set
+    self.tbvaldir = tbdir + '_val'
+    if not os.path.exists(self.tbvaldir):
+      os.makedirs(self.tbvaldir)
+    self.pretrained_model = pretrained_model
+
+  def snapshot(self, iter):
+
+    if not os.path.exists(self.output_dir):
+      os.makedirs(self.output_dir)
+
+    # Store the model snapshot
+    filename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(iter) + '.pth'
+    filename = os.path.join(self.output_dir, filename)
+    torch.save(self.net.state_dict(), filename)
+    print('Wrote snapshot to: {:s}'.format(filename))
+
+    # Also store some meta information, random state, etc.
+    nfilename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(iter) + '.pkl'
+    nfilename = os.path.join(self.output_dir, nfilename)
+    # current state of numpy random
+    st0 = np.random.get_state()
+    # current position in the database
+    cur = self.data_layer._cur
+    # current shuffled indexes of the database
+    perm = self.data_layer._perm
+    # current position in the validation database
+    cur_val = self.data_layer_val._cur
+    # current shuffled indexes of the validation database
+    perm_val = self.data_layer_val._perm
+
+    # Dump the meta info
+    with open(nfilename, 'wb') as fid:
+      pickle.dump(st0, fid, pickle.HIGHEST_PROTOCOL)
+      pickle.dump(cur, fid, pickle.HIGHEST_PROTOCOL)
+      pickle.dump(perm, fid, pickle.HIGHEST_PROTOCOL)
+      pickle.dump(cur_val, fid, pickle.HIGHEST_PROTOCOL)
+      pickle.dump(perm_val, fid, pickle.HIGHEST_PROTOCOL)
+      pickle.dump(iter, fid, pickle.HIGHEST_PROTOCOL)
+
+    return filename, nfilename
+
+  def from_snapshot(self, sfile, nfile):
+    print('Restoring model snapshots from {:s}'.format(sfile))
+    self.net.load_state_dict(torch.load(str(sfile)))
+    print('Restored.')
+    # Needs to restore the other hyper-parameters/states for training, (TODO xinlei) I have
+    # tried my best to find the random states so that it can be recovered exactly
+    # However the Tensorflow state is currently not available
+    with open(nfile, 'rb') as fid:
+      st0 = pickle.load(fid)
+      cur = pickle.load(fid)
+      perm = pickle.load(fid)
+      cur_val = pickle.load(fid)
+      perm_val = pickle.load(fid)
+      last_snapshot_iter = pickle.load(fid)
+
+      np.random.set_state(st0)
+      self.data_layer._cur = cur
+      self.data_layer._perm = perm
+      self.data_layer_val._cur = cur_val
+      self.data_layer_val._perm = perm_val
+
+    return last_snapshot_iter
+
+  def construct_graph(self):
+    # Set the random seed
+    torch.manual_seed(cfg.RNG_SEED)
+    # Build the main computation graph
+    self.net.create_architecture(self.imdb.num_classes, tag='default',
+                                            anchor_scales=cfg.ANCHOR_SCALES,
+                                            anchor_ratios=cfg.ANCHOR_RATIOS)
+    # Define the loss
+    # loss = layers['total_loss']
+    # Set learning rate and momentum
+    lr = cfg.TRAIN.LEARNING_RATE
+    params = []
+    for key, value in dict(self.net.named_parameters()).items():
+      if value.requires_grad:
+        if 'refine' in key and 'bias' in key:
+          params += [{'params':[value],'lr':10*lr*(cfg.TRAIN.DOUBLE_BIAS + 1), 'weight_decay': cfg.TRAIN.BIAS_DECAY and cfg.TRAIN.WEIGHT_DECAY or 0}]
+        elif 'refine' in key and 'bias' not in key:
+          params += [{'params':[value],'lr':10*lr, 'weight_decay': getattr(value, 'weight_decay', cfg.TRAIN.WEIGHT_DECAY)}]
+        elif 'refine' not in key and 'bias' in key:
+          params += [{'params':[value],'lr':lr*(cfg.TRAIN.DOUBLE_BIAS + 1), 'weight_decay': cfg.TRAIN.BIAS_DECAY and cfg.TRAIN.WEIGHT_DECAY or 0}]
+        else:
+          params += [{'params':[value],'lr':lr, 'weight_decay': getattr(value, 'weight_decay', cfg.TRAIN.WEIGHT_DECAY)}]
+    self.optimizer = torch.optim.SGD(params, momentum=cfg.TRAIN.MOMENTUM)
+    # Write the train and validation information to tensorboard
+    self.writer = tb.writer.FileWriter(self.tbdir)
+    self.valwriter = tb.writer.FileWriter(self.tbvaldir)
+
+    return lr, self.optimizer
+
+  def find_previous(self):
+    sfiles = os.path.join(self.output_dir, cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.pth')
+    sfiles = glob.glob(sfiles)
+    sfiles.sort(key=os.path.getmtime)
+    # Get the snapshot name in pytorch
+    redfiles = []
+    for stepsize in cfg.TRAIN.STEPSIZE:
+      redfiles.append(os.path.join(self.output_dir, 
+                      cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}.pth'.format(stepsize+1)))
+    #sfiles = [ss for ss in sfiles if ss not in redfiles]
+
+    nfiles = os.path.join(self.output_dir, cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.pkl')
+    nfiles = glob.glob(nfiles)
+    nfiles.sort(key=os.path.getmtime)
+    redfiles = [redfile.replace('.pth', '.pkl') for redfile in redfiles]
+    #nfiles = [nn for nn in nfiles if nn not in redfiles]
+
+    lsf = len(sfiles)
+    assert len(nfiles) == lsf
+
+    return lsf, nfiles, sfiles
+
+  def initialize(self):
+    # Initial file lists are empty
+    np_paths = []
+    ss_paths = []
+    # Fresh train directly from ImageNet weights
+    print('Loading initial model weights from {:s}'.format(self.pretrained_model))
+    self.net.load_pretrained_cnn(torch.load(self.pretrained_model))
+    print('Loaded.')
+    # Need to fix the variables before loading, so that the RGB weights are changed to BGR
+    # For VGG16 it also changes the convolutional weights fc6 and fc7 to
+    # fully connected weights
+    last_snapshot_iter = 0
+    lr = cfg.TRAIN.LEARNING_RATE
+    stepsizes = list(cfg.TRAIN.STEPSIZE)
+
+    return lr, last_snapshot_iter, stepsizes, np_paths, ss_paths
+
+  def restore(self, sfile, nfile):
+    # Get the most recent snapshot and restore
+    np_paths = [nfile]
+    ss_paths = [sfile]
+    # Restore model from snapshots
+    last_snapshot_iter = self.from_snapshot(sfile, nfile)
+    # Set the learning rate
+    lr_scale = 1
+    stepsizes = []
+    for stepsize in cfg.TRAIN.STEPSIZE:
+      if last_snapshot_iter > stepsize:
+        lr_scale *= cfg.TRAIN.GAMMA
+      else:
+        stepsizes.append(stepsize)
+    scale_lr(self.optimizer, lr_scale)
+    lr = cfg.TRAIN.LEARNING_RATE * lr_scale
+    return lr, last_snapshot_iter, stepsizes, np_paths, ss_paths
+
+  def remove_snapshot(self, np_paths, ss_paths):
+    to_remove = len(np_paths) - cfg.TRAIN.SNAPSHOT_KEPT
+    for c in range(to_remove):
+      nfile = np_paths[0]
+      os.remove(str(nfile))
+      np_paths.remove(nfile)
+
+    to_remove = len(ss_paths) - cfg.TRAIN.SNAPSHOT_KEPT
+    for c in range(to_remove):
+      sfile = ss_paths[0]
+      # To make the code compatible to earlier versions of Tensorflow,
+      # where the naming tradition for checkpoints are different
+      os.remove(str(sfile))
+      ss_paths.remove(sfile)
+
+  def train_model(self, max_iters):
+    # Build data layers for both training and validation set
+    self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
+    self.data_layer_val = RoIDataLayer(self.valroidb, self.imdb.num_classes, random=True)
+
+    # Construct the computation graph
+    lr, train_op = self.construct_graph()
+
+    # Find previous snapshots if there is any to restore from
+    lsf, nfiles, sfiles = self.find_previous()
+
+    # Initialize the variables or restore them from the last snapshot
+    if lsf == 0:
+      lr, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.initialize()
+    else:
+      lr, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.restore(str(sfiles[-1]), str(nfiles[-1]))
+    iter = last_snapshot_iter + 1
+    last_summary_time = time.time()
+    # Make sure the lists are not empty
+    stepsizes.append(max_iters)
+    stepsizes.reverse()
+    next_stepsize = stepsizes.pop()
+
+    self.net.train()
+    self.net.to(self.net._device)
+
+    while iter < max_iters + 1:
+      # Learning rate
+      if iter == next_stepsize + 1:
+        # Add snapshot here before reducing the learning rate
+        self.snapshot(iter)
+        lr *= cfg.TRAIN.GAMMA
+        scale_lr(self.optimizer, cfg.TRAIN.GAMMA)
+        next_stepsize = stepsizes.pop()
+      #if ((iter -1) % cfg.TRAIN.MIL_RECURRENT_STEP) == 0:
+      #  num_epoch = int((iter - 1) / cfg.TRAIN.MIL_RECURRENT_STEP) + 1
+      #  cfg.TRAIN.MIL_RECURRECT_WEIGHT = ((num_epoch - 1)/20.0)/1.5
+      #if iter == cfg.TRAIN.MIL_RECURRENT_STEP + 1:
+      #  cfg.TRAIN.MIL_RECURRECT_WEIGHT = cfg.TRAIN.MIL_RECURRECT_WEIGHT * 10
+    
+      utils.timer.timer.tic()
+      # Get training data, one batch at a time
+      blobs = self.data_layer.forward(iter)
+
+      now = time.time()
+      if iter == 1 or now - last_summary_time > cfg.TRAIN.SUMMARY_INTERVAL:
+        # Compute the graph with summary
+        cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, total_loss, summary = \
+          self.net.train_step_with_summary(blobs, self.optimizer, iter)
+        # cls_det_loss, refine_loss_1, refine_loss_2, total_loss, summary = \
+        #   self.net.train_step_with_summary(blobs, self.optimizer, iter)
+        #for _sum in summary: self.writer.add_summary(_sum, float(iter))
+        # Also check the summary on the validation set
+        #blobs_val = self.data_layer_val.forward()
+        #summary_val = self.net.get_summary(blobs_val, iter, drop_block)
+        #summary_val = self.net.get_summary(blobs_val, iter)
+
+        #for _sum in summary_val: self.valwriter.add_summary(_sum, float(iter))
+        last_summary_time = now
+      else:
+        # Compute the graph without summary
+        #cls_det_loss, refine_loss_1, refine_loss_2, total_loss = self.net.train_step(blobs, self.optimizer, iter)
+        cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, total_loss = self.net.train_step(blobs,self.optimizer,iter)
+      utils.timer.timer.toc()
+
+      # Display training information
+      if iter % (cfg.TRAIN.DISPLAY) == 0:
+        # print('iter: %d / %d, total loss: %.6f\n >>> cls_det_loss: %.6f\n '
+        #       '>>> refine_loss_1: %.6f\n >>> refine_loss_2: %.6f\n >>> lr: %f' % \
+        #       (iter, max_iters, total_loss, cls_det_loss, refine_loss_1, refine_loss_2, lr))
+        print('iter: %d / %d, total loss: %.6f\n >>> cls_det_loss: %.6f\n '
+              '>>> refine_loss_1: %.6f\n >>> refine_loss_2: %.6f\n >>> consistency_loss: %.6f\n >>> lr: %f' % \
+              (iter, max_iters, total_loss, cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, lr))
+        print('speed: {:.3f}s / iter'.format(utils.timer.timer.average_time()))
+
+        # for k in utils.timer.timer._average_time.keys():
+        #   print(k, utils.timer.timer.average_time(k))
+
+      # Snapshotting
+      if iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
+        last_snapshot_iter = iter
+        ss_path, np_path = self.snapshot(iter)
+        np_paths.append(np_path)
+        ss_paths.append(ss_path)
+
+        # Remove the old snapshots if there are too many
+        if len(np_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
+          self.remove_snapshot(np_paths, ss_paths)
+
+      iter += 1
+
+    if last_snapshot_iter != iter - 1:
+      self.snapshot(iter - 1)
+
+    self.writer.close()
+    self.valwriter.close()
+
+
+def get_training_roidb(imdb):
+  """Returns a roidb (Region of Interest database) for use in training."""
+  if cfg.TRAIN.USE_FLIPPED:
+    print('Appending horizontally-flipped training examples...')
+    imdb.append_flipped_images()
+    print('done')
+
+  print('Preparing training data...')
+  rdl_roidb.prepare_roidb(imdb)
+  print('done')
+
+  return imdb.roidb
+
+
+def filter_roidb(roidb):
+  """Remove roidb entries that have no usable RoIs."""
+
+  def is_valid(entry):
+    # Valid images have:
+    #   (1) At least one foreground RoI OR
+    #   (2) At least one background RoI
+    overlaps = entry['max_overlaps']
+    # find boxes with sufficient overlap
+    fg_inds = np.where(overlaps >= cfg.TRAIN.FG_THRESH)[0]
+    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
+    bg_inds = np.where((overlaps < cfg.TRAIN.BG_THRESH_HI) &
+                       (overlaps >= cfg.TRAIN.BG_THRESH_LO))[0]
+    # image is only valid if such boxes exist
+    valid = len(fg_inds) > 0 or len(bg_inds) > 0
+    return valid
+
+  num = len(roidb)
+  filtered_roidb = [entry for entry in roidb if is_valid(entry)]
+  num_after = len(filtered_roidb)
+  print('Filtered {} roidb entries: {} -> {}'.format(num - num_after,
+                                                     num, num_after))
+  return filtered_roidb
+
+
+def train_net(network, imdb, roidb, valroidb, output_dir, tb_dir,
+              pretrained_model=None,
+              max_iters=40000):
+  """Train a Faster R-CNN network."""
+  roidb = filter_roidb(roidb)
+  valroidb = filter_roidb(valroidb)
+
+  sw = SolverWrapper(network, imdb, roidb, valroidb, output_dir, tb_dir,
+                     pretrained_model=pretrained_model)
+
+  print('Solving...')
+  sw.train_model(max_iters)
+  print('done solving')
diff --git a/lib/nets/__init__.py b/lib/nets/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/lib/nets/__pycache__/__init__.cpython-36.pyc b/lib/nets/__pycache__/__init__.cpython-36.pyc
new file mode 100644
index 0000000..4ef374d
Binary files /dev/null and b/lib/nets/__pycache__/__init__.cpython-36.pyc differ
diff --git a/lib/nets/__pycache__/__init__.cpython-37.pyc b/lib/nets/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..dcc2a9c
Binary files /dev/null and b/lib/nets/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/nets/__pycache__/mobilenet_v1.cpython-36.pyc b/lib/nets/__pycache__/mobilenet_v1.cpython-36.pyc
new file mode 100644
index 0000000..cb69272
Binary files /dev/null and b/lib/nets/__pycache__/mobilenet_v1.cpython-36.pyc differ
diff --git a/lib/nets/__pycache__/mobilenet_v1.cpython-37.pyc b/lib/nets/__pycache__/mobilenet_v1.cpython-37.pyc
new file mode 100644
index 0000000..ff4340b
Binary files /dev/null and b/lib/nets/__pycache__/mobilenet_v1.cpython-37.pyc differ
diff --git a/lib/nets/__pycache__/network.cpython-36.pyc b/lib/nets/__pycache__/network.cpython-36.pyc
new file mode 100644
index 0000000..f70d3be
Binary files /dev/null and b/lib/nets/__pycache__/network.cpython-36.pyc differ
diff --git a/lib/nets/__pycache__/network.cpython-37.pyc b/lib/nets/__pycache__/network.cpython-37.pyc
new file mode 100644
index 0000000..359bea7
Binary files /dev/null and b/lib/nets/__pycache__/network.cpython-37.pyc differ
diff --git a/lib/nets/__pycache__/resnet_v1.cpython-36.pyc b/lib/nets/__pycache__/resnet_v1.cpython-36.pyc
new file mode 100644
index 0000000..8be4e6f
Binary files /dev/null and b/lib/nets/__pycache__/resnet_v1.cpython-36.pyc differ
diff --git a/lib/nets/__pycache__/resnet_v1.cpython-37.pyc b/lib/nets/__pycache__/resnet_v1.cpython-37.pyc
new file mode 100644
index 0000000..d54ab65
Binary files /dev/null and b/lib/nets/__pycache__/resnet_v1.cpython-37.pyc differ
diff --git a/lib/nets/__pycache__/vgg16.cpython-36.pyc b/lib/nets/__pycache__/vgg16.cpython-36.pyc
new file mode 100644
index 0000000..bd553d8
Binary files /dev/null and b/lib/nets/__pycache__/vgg16.cpython-36.pyc differ
diff --git a/lib/nets/__pycache__/vgg16.cpython-37.pyc b/lib/nets/__pycache__/vgg16.cpython-37.pyc
new file mode 100644
index 0000000..2bae43d
Binary files /dev/null and b/lib/nets/__pycache__/vgg16.cpython-37.pyc differ
diff --git a/lib/nets/mobilenet_v1.py b/lib/nets/mobilenet_v1.py
new file mode 100644
index 0000000..d68b408
--- /dev/null
+++ b/lib/nets/mobilenet_v1.py
@@ -0,0 +1,262 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+import numpy as np
+from collections import namedtuple, OrderedDict
+
+from nets.network import Network
+from model.config import cfg
+
+# The following is adapted from:
+# https://github.com/tensorflow/models/blob/master/slim/nets/mobilenet_v1.py
+
+# Conv and DepthSepConv named tuple define layers of the MobileNet architecture
+# Conv defines 3x3 convolution layers
+# DepthSepConv defines 3x3 depthwise convolution followed by 1x1 convolution.
+# stride is the stride of the convolution
+# depth is the number of channels or filters in a layer
+Conv = namedtuple('Conv', ['kernel', 'stride', 'depth'])
+DepthSepConv = namedtuple('DepthSepConv', ['kernel', 'stride', 'depth'])
+
+# _CONV_DEFS specifies the MobileNet body
+_CONV_DEFS = [
+    Conv(kernel=3, stride=2, depth=32),
+    DepthSepConv(kernel=3, stride=1, depth=64),
+    DepthSepConv(kernel=3, stride=2, depth=128),
+    DepthSepConv(kernel=3, stride=1, depth=128),
+    DepthSepConv(kernel=3, stride=2, depth=256),
+    DepthSepConv(kernel=3, stride=1, depth=256),
+    DepthSepConv(kernel=3, stride=2, depth=512),
+    DepthSepConv(kernel=3, stride=1, depth=512),
+    DepthSepConv(kernel=3, stride=1, depth=512),
+    DepthSepConv(kernel=3, stride=1, depth=512),
+    DepthSepConv(kernel=3, stride=1, depth=512),
+    DepthSepConv(kernel=3, stride=1, depth=512),
+    # use stride 1 for the 13th layer
+    DepthSepConv(kernel=3, stride=1, depth=1024),
+    DepthSepConv(kernel=3, stride=1, depth=1024)
+]
+
+def mobilenet_v1_base(final_endpoint='Conv2d_13_pointwise',
+                      min_depth=8,
+                      depth_multiplier=1.0,
+                      conv_defs=None,
+                      output_stride=None):
+    """Mobilenet v1.
+
+    Constructs a Mobilenet v1 network from inputs to the given final endpoint.
+
+    Args:
+        inputs: a tensor of shape [batch_size, height, width, channels].
+        final_endpoint: specifies the endpoint to construct the network up to. It
+            can be one of ['Conv2d_0', 'Conv2d_1_pointwise', 'Conv2d_2_pointwise',
+            'Conv2d_3_pointwise', 'Conv2d_4_pointwise', 'Conv2d_5_pointwise',
+            'Conv2d_6_pointwise', 'Conv2d_7_pointwise', 'Conv2d_8_pointwise',
+            'Conv2d_9_pointwise', 'Conv2d_10_pointwise', 'Conv2d_11_pointwise',
+            'Conv2d_12_pointwise', 'Conv2d_13_pointwise'].
+        min_depth: Minimum depth value (number of channels) for all convolution ops.
+            Enforced when depth_multiplier < 1, and not an active constraint when
+            depth_multiplier >= 1.
+        depth_multiplier: Float multiplier for the depth (number of channels)
+            for all convolution ops. The value must be greater than zero. Typical
+            usage will be to set this value in (0, 1) to reduce the number of
+            parameters or computation cost of the model.
+        conv_defs: A list of ConvDef namedtuples specifying the net architecture.
+        output_stride: An integer that specifies the requested ratio of input to
+            output spatial resolution. If not None, then we invoke atrous convolution
+            if necessary to prevent the network from reducing the spatial resolution
+            of the activation maps. Allowed values are 8 (accurate fully convolutional
+            mode), 16 (fast fully convolutional mode), 32 (classification mode).
+        scope: Optional variable_scope.
+
+    Returns:
+        tensor_out: output tensor corresponding to the final_endpoint.
+        end_points: a set of activations for external use, for example summaries or
+                                losses.
+
+    Raises:
+        ValueError: if final_endpoint is not set to one of the predefined values,
+                                or depth_multiplier <= 0, or the target output_stride is not
+                                allowed.
+    """
+    depth = lambda d: max(int(d * depth_multiplier), min_depth)
+    end_points = OrderedDict()
+
+    # Used to find thinned depths for each layer.
+    if depth_multiplier <= 0:
+        raise ValueError('depth_multiplier is not greater than zero.')
+
+    if conv_defs is None:
+        conv_defs = _CONV_DEFS
+
+    if output_stride is not None and output_stride not in [8, 16, 32]:
+        raise ValueError('Only allowed output_stride values are 8, 16, 32.')
+
+    def conv_bn(in_channels, out_channels, kernel_size=3, stride=1):
+        return nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size, stride, (kernel_size - 1) // 2, bias=False),
+            nn.BatchNorm2d(out_channels),
+            nn.ReLU6(inplace=True)
+        )
+
+    def conv_dw(in_channels, kernel_size=3, stride=1, dilation=1):
+        return nn.Sequential(
+            nn.Conv2d(in_channels, in_channels, kernel_size, stride, (kernel_size - 1) // 2,\
+                      groups=in_channels, dilation=dilation, bias=False),
+            nn.BatchNorm2d(in_channels),
+            nn.ReLU6(inplace=True)
+        )
+
+    def conv_pw(in_channels, out_channels, kernel_size=3, stride=1, dilation=1):
+        return nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size, stride, 0, bias=False),
+            nn.BatchNorm2d(out_channels),
+            nn.ReLU6(inplace=True),
+        )
+
+    # The current_stride variable keeps track of the output stride of the
+    # activations, i.e., the running product of convolution strides up to the
+    # current network layer. This allows us to invoke atrous convolution
+    # whenever applying the next convolution would result in the activations
+    # having output stride larger than the target output_stride.
+    current_stride = 1
+
+    # The atrous convolution rate parameter.
+    rate = 1
+
+    in_channels = 3
+    for i, conv_def in enumerate(conv_defs):
+        end_point_base = 'Conv2d_%d' % i
+
+        if output_stride is not None and current_stride == output_stride:
+            # If we have reached the target output_stride, then we need to employ
+            # atrous convolution with stride=1 and multiply the atrous rate by the
+            # current unit's stride for use in subsequent layers.
+            layer_stride = 1
+            layer_rate = rate
+            rate *= conv_def.stride
+        else:
+            layer_stride = conv_def.stride
+            layer_rate = 1
+            current_stride *= conv_def.stride
+
+        out_channels = depth(conv_def.depth)
+        if isinstance(conv_def, Conv):
+            end_point = end_point_base
+            end_points[end_point] = conv_bn(in_channels, out_channels, conv_def.kernel,
+                                            stride=conv_def.stride)
+            if end_point == final_endpoint:
+                return nn.Sequential(end_points)
+
+        elif isinstance(conv_def, DepthSepConv):
+            end_points[end_point_base] = nn.Sequential(OrderedDict([
+                ('depthwise', conv_dw(in_channels, conv_def.kernel, stride=layer_stride, dilation=layer_rate)),
+                ('pointwise', conv_pw(in_channels, out_channels, 1, stride=1))]))
+
+            if end_point_base + '_pointwise' == final_endpoint:
+                return nn.Sequential(end_points)
+        else:
+            raise ValueError('Unknown convolution type %s for layer %d'
+                                                % (conv_def.ltype, i))
+        in_channels = out_channels
+    raise ValueError('Unknown final endpoint %s' % final_endpoint)
+
+class mobilenetv1(Network):
+  def __init__(self):
+    Network.__init__(self)
+    self._feat_stride = [16, ]
+    self._feat_compress = [1. / float(self._feat_stride[0]), ]
+    self._depth_multiplier = cfg.MOBILENET.DEPTH_MULTIPLIER
+    self._net_conv_channels = 512
+    self._fc7_channels = 1024
+
+  def init_weights(self):
+    def normal_init(m, mean, stddev, truncated=False):
+      """
+      weight initalizer: truncated normal and random normal.
+      """
+      if m.__class__.__name__.find('Conv') == -1:
+        return
+      if truncated:
+        m.weight.data.normal_().fmod_(2).mul_(stddev).add_(mean) # not a perfect approximation
+      else:
+        m.weight.data.normal_(mean, stddev)
+      if m.bias is not None: m.bias.data.zero_()
+
+    self.mobilenet.apply(lambda m: normal_init(m, 0, 0.09, True))
+    normal_init(self.rpn_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+    normal_init(self.rpn_cls_score_net, 0, 0.01,  cfg.TRAIN.TRUNCATED)
+    normal_init(self.rpn_bbox_pred_net, 0, 0.01,  cfg.TRAIN.TRUNCATED)
+    normal_init(self.cls_score_net, 0, 0.01,  cfg.TRAIN.TRUNCATED)
+    normal_init(self.bbox_pred_net, 0, 0.001,  cfg.TRAIN.TRUNCATED)
+
+  def _image_to_head(self):
+    net_conv = self._layers['head'](self._image)
+    self._act_summaries['conv'] = net_conv
+
+    return net_conv
+
+  def _head_to_tail(self, pool5):
+    fc7 = self._layers['tail'](pool5)
+    fc7 = fc7.mean(3).mean(2)
+    return fc7
+
+  def _init_head_tail(self):
+    self.mobilenet = mobilenet_v1_base()
+
+    # Fix blocks
+    assert (0 <= cfg.MOBILENET.FIXED_LAYERS <= 12)
+    for m in list(self.mobilenet.children())[:cfg.MOBILENET.FIXED_LAYERS]:
+      for p in m.parameters():
+        p.requires_grad = False
+
+    def set_bn_fix(m):
+      classname = m.__class__.__name__
+      if classname.find('BatchNorm') != -1:
+        for p in m.parameters(): p.requires_grad=False
+
+    self.mobilenet.apply(set_bn_fix)
+
+    # Add weight decay
+    def l2_regularizer(m, wd, regu_depth):
+      if m.__class__.__name__.find('Conv') != -1:
+        if regu_depth or m.groups == 1:
+            m.weight.weight_decay = wd
+        else:
+            m.weight.weight_decay = 0
+    self.mobilenet.apply(lambda x: l2_regularizer(x, cfg.MOBILENET.WEIGHT_DECAY, cfg.MOBILENET.REGU_DEPTH))
+
+    # Build mobilenet.
+    self._layers['head'] = nn.Sequential(*list(self.mobilenet.children())[:12])
+    self._layers['tail'] = nn.Sequential(*list(self.mobilenet.children())[12:])
+
+  def train(self, mode=True):
+    # Override train so that the training mode is set as we want
+    nn.Module.train(self, mode)
+    if mode:
+      # Set fixed blocks to be in eval mode (not really doing anything)
+      for m in list(self.mobilenet.children())[:cfg.MOBILENET.FIXED_LAYERS]:
+        m.eval()
+
+      # Set batchnorm always in eval mode during training
+      def set_bn_eval(m):
+        classname = m.__class__.__name__
+        if classname.find('BatchNorm') != -1:
+          m.eval()
+
+      self.mobilenet.apply(set_bn_eval)
+
+  def load_pretrained_cnn(self, state_dict):
+    print('Warning: No available pretrained model yet')
+    self.mobilenet.load_state_dict({k: state_dict['features.'+k] for k in list(self.mobilenet.state_dict())})
diff --git a/lib/nets/network.py b/lib/nets/network.py
new file mode 100644
index 0000000..c24f999
--- /dev/null
+++ b/lib/nets/network.py
@@ -0,0 +1,871 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from utils.visualization import draw_bounding_boxes
+
+from ops.roi_pool import RoIPool
+from ops.roi_align import RoIAlign
+from ops.roi_ring_pool import RoIRingPool
+
+from model.config import cfg
+from utils.bbox import bbox_overlaps
+import tensorboardX as tb
+
+from scipy.misc import imresize
+from sklearn.cluster import KMeans
+import random
+from torch.autograd import Variable
+
+import numpy.random as npr
+
+import math
+
+try:
+    xrange  # Python 2
+except NameError:
+    xrange = range  # Python 3
+
+
+class Network(nn.Module):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self._predictions = {}
+        self._losses = {}
+        self._layers = {}
+        self._gt_image = None
+        self._event_summaries = {}
+        self._image_gt_summaries = {}
+        self._device = 'cuda'
+        self.RoIPool = RoIPool((cfg.POOLING_SIZE, cfg.POOLING_SIZE), 1. / 16)
+        self.RoIAlign = RoIAlign((cfg.POOLING_SIZE, cfg.POOLING_SIZE), 1. / 16)
+        self.RoIRingPool = RoIRingPool(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1. / 16, 0., 1.0)
+        self.RoIRingPool_context = RoIRingPool(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1. / 16, 1.0, 1.8)
+        self.RoIRingPool_frame = RoIRingPool(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1. / 16, 1.0 / 1.8, 1.0)
+        self.aug_time = 4
+        self.ca_iw = True
+
+    def _add_gt_image(self):
+        # add back mean
+        image = self._image_gt_summaries['image'] + cfg.PIXEL_MEANS
+        image = imresize(image[0], self._im_info[:2] / self._im_info[2])
+        # BGR to RGB (opencv uses BGR)
+        self._gt_image = image[np.newaxis, :, :, ::-1].copy(order='C')
+
+    def _add_gt_image_summary(self):
+        # use a customized visualization function to visualize the boxes
+        self._add_gt_image()
+        image = draw_bounding_boxes(self._gt_image, self._image_gt_summaries['gt_boxes'], self._image_gt_summaries['im_info'])
+        return tb.summary.image('GROUND_TRUTH', image[0].astype('float32').swapaxes(1, 0).swapaxes(2, 0) / 255.0)
+
+    def _inverted_attention(self, bbox_feats_new, gt, keep_inds, refine_branch, step, fg_num, bg_num):
+        if step <= (cfg.TRAIN.STEP_ITERS + 10000):
+            fg_drop_per = 6
+            bg_drop_per = fg_drop_per * 3
+            th_l = 25
+            th_s = 16
+        else:
+            fg_drop_per = 5
+            bg_drop_per = fg_drop_per * 3
+            th_l = 29
+            th_s = 20
+
+        self.eval()
+        bbox_feats_new = Variable(bbox_feats_new.data, requires_grad=True)
+        bbox_feats_new_new = self._head_to_tail(bbox_feats_new)
+        if refine_branch == 1:
+            output_score = self.refine_net_1(bbox_feats_new_new)
+        elif refine_branch == 2:
+            output_score = self.refine_net_2(bbox_feats_new_new)
+        else:
+            print('no refine branch')
+        class_num = output_score.shape[1]
+        index = gt
+        num_rois = bbox_feats_new.shape[0]
+        num_channel = bbox_feats_new.shape[1]
+        one_hot = torch.zeros((1), dtype=torch.float32).cuda()
+        one_hot = Variable(one_hot, requires_grad=False)
+        sp_i = torch.ones([2, num_rois]).long()
+        sp_i[0, :] = torch.arange(num_rois)
+        sp_i[1, :] = torch.from_numpy(index)
+        sp_v = torch.ones([num_rois])
+        one_hot_sparse = torch.sparse.FloatTensor(sp_i, sp_v, torch.Size([num_rois, class_num])).to_dense().cuda()
+        one_hot_sparse = Variable(one_hot_sparse, requires_grad=False)
+        one_hot = torch.sum(output_score * one_hot_sparse)
+        self.vgg.classifier.zero_grad()
+        if refine_branch == 1:
+            self.refine_net_1.zero_grad()
+        elif refine_branch == 2:
+            self.refine_net_2.zero_grad()
+        else:
+            print('no refine branch')
+        one_hot.backward()
+        grads_val = bbox_feats_new.grad.clone().detach()
+        grad_channel_mean = torch.mean(grads_val.view(num_rois, num_channel, -1), dim=2)
+        grad_channel_mean = grad_channel_mean.view(num_rois, num_channel, 1, 1)
+        att_all = torch.sum(bbox_feats_new * grad_channel_mean, 1)
+        att_all = att_all.view(num_rois, 49)
+
+        self.vgg.classifier.zero_grad()
+        if refine_branch == 1:
+            self.refine_net_1.zero_grad()
+        elif refine_branch == 2:
+            self.refine_net_2.zero_grad()
+        else:
+            print('no refine branch')
+
+        thl_mask_value = torch.sort(att_all, dim=1, descending=True)[0][:, th_l]
+        thl_mask_value = thl_mask_value.view(num_rois, 1).expand(num_rois, 49)
+        mask_all_cuda = torch.where(att_all > thl_mask_value, torch.zeros(att_all.shape).cuda(), torch.ones(att_all.shape).cuda())
+        mask_all = mask_all_cuda.detach().cpu().numpy()
+        mask_all_new = np.ones((num_rois, 49), dtype=np.float32)
+        for q in keep_inds:
+            mask_all_temp = np.ones((49), dtype=np.float32)
+            zero_index = np.where(mask_all[q, :] == 0)[0]
+            num_zero_index = zero_index.size
+            if num_zero_index >= th_s:
+                dumy_index = npr.choice(zero_index, size=th_s, replace=False)
+            else:
+                zero_index = np.arange(49)
+                dumy_index = npr.choice(zero_index, size=th_s, replace=False)
+            mask_all_temp[dumy_index] = 0
+            mask_all_new[q, :] = mask_all_temp
+        mask_all = torch.from_numpy(mask_all_new.reshape(num_rois, 7, 7)).cuda()
+        mask_all = mask_all.view(num_rois, 1, 7, 7)
+
+        pooled_feat_before_after = torch.cat((bbox_feats_new, bbox_feats_new * mask_all), dim=0)
+        pooled_feat_before_after = self._head_to_tail(pooled_feat_before_after)
+        if refine_branch == 1:
+            cls_score_before_after = self.refine_net_1(pooled_feat_before_after)
+        elif refine_branch == 2:
+            cls_score_before_after = self.refine_net_2(pooled_feat_before_after)
+        else:
+            print('no refine branch')
+        cls_prob_before_after = F.softmax(cls_score_before_after, dim=1)
+        class_num = cls_prob_before_after.shape[1]
+        cls_prob_before = cls_prob_before_after[0: num_rois]
+        cls_prob_after = cls_prob_before_after[num_rois: num_rois * 2]
+        label_gt = torch.from_numpy(gt).cuda()
+        prepare_mask_fg_num = fg_num
+        prepare_mask_bg_num = bg_num
+        sp_i = torch.ones([2, num_rois]).long()
+        sp_i[0, :] = torch.arange(num_rois)
+        sp_i[1, :] = label_gt
+        sp_v = torch.ones([num_rois])
+        one_hot_sparse = torch.sparse.FloatTensor(sp_i, sp_v, torch.Size([num_rois, class_num])).to_dense().cuda()
+        before_vector = torch.sum(one_hot_sparse * cls_prob_before, dim=1)
+        after_vector = torch.sum(one_hot_sparse * cls_prob_after, dim=1)
+        change_vector = before_vector - after_vector - 0.02
+        change_vector = torch.where(change_vector > 0, change_vector, torch.zeros(change_vector.shape).cuda())
+        fg_index = torch.where(label_gt > 0, torch.ones(before_vector.shape).cuda(), torch.zeros(before_vector.shape).cuda())
+        bg_index = 1 - fg_index
+
+        if fg_index.nonzero().shape[0] != 0:
+            not_01_fg_index = fg_index.nonzero()[:, 0].long()
+        else:
+            not_01_fg_index = torch.tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]).cuda().long()  # for corner case
+        not_01_bg_index = bg_index.nonzero()[:, 0].long()
+        change_vector_fg = change_vector[not_01_fg_index]
+        change_vector_bg = change_vector[not_01_bg_index]
+        for_fg_change_vector = change_vector.clone()
+        for_bg_change_vector = change_vector.clone()
+        for_fg_change_vector[not_01_bg_index] = -10000
+        for_bg_change_vector[not_01_fg_index] = -10000
+
+        th_fg_value = torch.sort(change_vector_fg, dim=0, descending=True)[0][int(round(float(prepare_mask_fg_num) / fg_drop_per))]
+        drop_index_fg = for_fg_change_vector.gt(th_fg_value)
+        th_bg_value = torch.sort(change_vector_bg, dim=0, descending=True)[0][int(round(float(prepare_mask_bg_num) / bg_drop_per))]
+        drop_index_bg = for_bg_change_vector.gt(th_bg_value)
+        drop_index_fg_bg = drop_index_fg + drop_index_bg
+        ignore_index_fg_bg = 1 - drop_index_fg_bg
+        not_01_ignore_index_fg_bg = ignore_index_fg_bg.nonzero()[:, 0]
+        mask_all[not_01_ignore_index_fg_bg.long(), :] = 1
+        self.train()
+        return mask_all
+
+    def _normalize_atten_maps(self, atten_maps):
+        atten_shape = atten_maps.size()
+        batch_mins, _ = torch.min(atten_maps.view(atten_shape[0:-2] + (-1,)), dim=-1, keepdim=True)
+        batch_maxs, _ = torch.max(atten_maps.view(atten_shape[0:-2] + (-1,)), dim=-1, keepdim=True)
+        atten_normed = torch.div(atten_maps.view(atten_shape[0:-2] + (-1,))-batch_mins,
+                                 batch_maxs - batch_mins + 1e-7)
+        atten_normed = atten_normed.view(atten_shape)
+        return atten_normed
+
+    def _rampweight(self, iteration):
+        ramp_up_end = 45000
+        ramp_down_start = 100000
+        if (iteration < ramp_up_end):
+            ramp_weight = math.exp(-5 * math.pow((1 - iteration / ramp_up_end), 2))
+        elif (iteration > ramp_down_start):
+            ramp_weight = math.exp(-12.5 * math.pow((1 - (100000 - iteration) / 20000), 2))
+        else:
+            ramp_weight = 1
+
+        if (iteration <= 45000):
+            ramp_weight = 0
+            self.ca_iw = False
+        return ramp_weight
+
+    def _add_losses(self, roi_labels_1, keep_inds_1, roi_labels_2, keep_inds_2, step=None, rois=None):
+        det_cls_prob = self._predictions['det_cls_prob']
+        det_cls_prob = det_cls_prob.view(-1)
+        label = self._image_level_label.view(-1)
+        pi = self.ss_boxes_indexes.shape[0]
+        rampweight = self._rampweight(step)
+
+        refine_prob_1 = self._predictions['refine_prob_1']
+        refine_prob_2 = self._predictions['refine_prob_2']
+        # refine_prob_3 = self._predictions['refine_prob_3']
+
+        # caculating the loss of the first branch
+        roi_labels, keep_inds = roi_labels_1, keep_inds_1,
+        roi_labels_each = torch.tensor(roi_labels[0][keep_inds[0], :], dtype=torch.float32).cuda()
+        refine_loss_1 = - torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_1[keep_inds[0]]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[1][keep_inds[1], :], dtype=torch.float32).cuda()
+        refine_loss_1 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_1[keep_inds[1] + pi]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[2][keep_inds[2], :], dtype=torch.float32).cuda()
+        refine_loss_1 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_1[keep_inds[2] + pi * 2]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[3][keep_inds[3], :], dtype=torch.float32).cuda()
+        refine_loss_1 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_1[keep_inds[3] + pi * 3]))) / roi_labels_each.shape[0]
+
+        consistency_conf_loss = 0
+        if self.ca_iw:
+            keep_inds_new = np.concatenate((keep_inds[0], keep_inds[0]+pi, keep_inds[0]+pi*2, keep_inds[0]+pi*3))
+            num_each = int(keep_inds_new.shape[0] / 4)
+            rois_self_attention_1 = torch.mean(rois[keep_inds_new], dim=1)
+            rois_self_attention_1 = torch.sigmoid(self._normalize_atten_maps(rois_self_attention_1))
+
+            rois_self_attention_gt_1 = rois_self_attention_1.clone().detach()
+            rois_self_attention_gt_1_1 = torch.max(rois_self_attention_gt_1[0:num_each], rois_self_attention_gt_1[num_each:num_each * 2].flip(dims=[2]))
+            rois_self_attention_gt_1_2 = torch.max(rois_self_attention_gt_1[num_each * 2:num_each * 3], rois_self_attention_gt_1[num_each * 3:num_each * 4].flip(dims=[2]))
+            rois_self_attention_gt_1 = torch.max(rois_self_attention_gt_1_1, rois_self_attention_gt_1_2)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_1[0:num_each], rois_self_attention_gt_1)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_1[num_each:num_each * 2], rois_self_attention_gt_1.flip(dims=[2]))
+            consistency_conf_loss += F.mse_loss(rois_self_attention_1[num_each * 2:num_each * 3], rois_self_attention_gt_1)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_1[num_each * 3:num_each * 4], rois_self_attention_gt_1.flip(dims=[2]))
+
+        # caculating the loss of the second branch
+        roi_labels, keep_inds = roi_labels_2, keep_inds_2
+        roi_labels_each = torch.tensor(roi_labels[0][keep_inds[0], :], dtype=torch.float32).cuda()
+        refine_loss_2 = - torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_2[keep_inds[0]]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[1][keep_inds[1], :], dtype=torch.float32).cuda()
+        refine_loss_2 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_2[keep_inds[1] + pi]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[2][keep_inds[2], :], dtype=torch.float32).cuda()
+        refine_loss_2 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_2[keep_inds[2] + pi * 2]))) / roi_labels_each.shape[0]
+        roi_labels_each = torch.tensor(roi_labels[3][keep_inds[3], :], dtype=torch.float32).cuda()
+        refine_loss_2 -= torch.sum(torch.mul(roi_labels_each, torch.log(refine_prob_2[keep_inds[3] + pi * 3]))) / roi_labels_each.shape[0]
+
+        if self.ca_iw:
+            keep_inds_new = np.concatenate((keep_inds[0], keep_inds[0]+pi, keep_inds[0]+pi*2, keep_inds[0]+pi*3))
+            num_each = int(keep_inds_new.shape[0] / 4)
+            rois_self_attention_2 = torch.mean(rois[keep_inds_new], dim=1)
+            rois_self_attention_2 = torch.sigmoid(self._normalize_atten_maps(rois_self_attention_2))
+
+            rois_self_attention_gt_2 = rois_self_attention_2.clone().detach()
+            rois_self_attention_gt_2_1 = torch.max(rois_self_attention_gt_2[0:num_each], rois_self_attention_gt_2[num_each:num_each*2].flip(dims=[2]))
+            rois_self_attention_gt_2_2 = torch.max(rois_self_attention_gt_2[num_each*2:num_each*3], rois_self_attention_gt_2[num_each*3:num_each*4].flip(dims=[2]))
+            rois_self_attention_gt_2 = torch.max(rois_self_attention_gt_2_1, rois_self_attention_gt_2_2)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_2[0:num_each], rois_self_attention_gt_2)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_2[num_each:num_each*2], rois_self_attention_gt_2.flip(dims=[2]))
+            consistency_conf_loss += F.mse_loss(rois_self_attention_2[num_each*2:num_each*3], rois_self_attention_gt_2)
+            consistency_conf_loss += F.mse_loss(rois_self_attention_2[num_each*3:num_each*4], rois_self_attention_gt_2.flip(dims=[2]))
+            consistency_conf_loss /= num_each
+
+        label_new = torch.cat((label, label, label, label))
+        label_new = label_new.clone().detach().float().to(det_cls_prob.device)
+        zeros = torch.zeros(det_cls_prob.shape, dtype=det_cls_prob.dtype, device=det_cls_prob.device)
+        max_zeros = torch.max(zeros, 1 - torch.mul(label_new, det_cls_prob))
+        cls_det_loss = torch.sum(max_zeros)
+        loss = cls_det_loss / 20 + refine_loss_1 * 0.1 + refine_loss_2 * 0.1 + consistency_conf_loss * 0.1
+        loss /= float(self.aug_time)
+
+        self._losses['total_loss'] = loss
+        self._losses['cls_det_loss'] = cls_det_loss / 20
+        self._losses['refine_loss_1'] = refine_loss_1 * 0.1
+        self._losses['refine_loss_2'] = refine_loss_2 * 0.1
+        # self._losses['refine_loss_3'] = refine_loss_3
+        if self.ca_iw is False:
+            consistency_conf_loss = torch.zeros([1])
+        self._losses['consistency_loss'] = consistency_conf_loss
+        for k in self._losses.keys():
+            self._event_summaries[k] = self._losses[k]
+        return loss
+
+    def _region_classification_test(self, fc7_roi, fc7_context, fc7_frame):
+        refine_score_1 = self.refine_net_1(fc7_roi)
+        refine_score_2 = self.refine_net_2(fc7_roi)
+        # refine_score_3 = self.refine_net_3(fc7_roi)
+        cls_score = self.cls_score_net(fc7_roi)
+        context_score = self.det_score_net(fc7_context)
+        frame_score = self.det_score_net(fc7_frame)
+        det_score = frame_score - context_score
+
+        cls_prob = F.softmax(cls_score, dim=1)
+        det_prob = F.softmax(det_score, dim=0)
+        refine_prob_1 = F.softmax(refine_score_1, dim=1)
+        refine_prob_2 = F.softmax(refine_score_2, dim=1)
+        # refine_prob_3 = F.softmax(refine_score_3, dim=1)
+
+        det_cls_prob_product = torch.mul(cls_score, det_prob)
+        det_cls_prob = torch.sum(det_cls_prob_product, 0)
+        # bbox_pred = self.bbox_pred_net(fc7)
+        bbox_pred = torch.zeros(cls_prob.shape[0], 80)
+
+        self._predictions['refine_prob_1'] = refine_prob_1
+        self._predictions['refine_prob_2'] = refine_prob_2
+        # self._predictions['refine_prob_3'] = refine_prob_3
+        self._predictions["bbox_pred"] = bbox_pred
+        self._predictions['det_cls_prob_product'] = det_cls_prob_product
+        self._predictions['det_cls_prob'] = det_cls_prob
+        return cls_prob, det_prob, bbox_pred, det_cls_prob_product, det_cls_prob
+
+    def _region_classification_train(self, pool5_roi, fc7_roi, fc7_context, fc7_frame, step):
+        bbox_feats_new = pool5_roi.clone().detach()
+        fc7_roi_new = fc7_roi.clone().detach()
+        refine_score_1_new = self.refine_net_1(fc7_roi_new)
+        refine_prob_1_new = F.softmax(refine_score_1_new, dim=1)
+
+        cls_score = self.cls_score_net(fc7_roi)
+        context_score = self.det_score_net(fc7_context)
+        frame_score = self.det_score_net(fc7_frame)
+        det_score = frame_score - context_score
+
+        cls_prob = F.softmax(cls_score, dim=1)
+        pi = self.ss_boxes_indexes.shape[0]
+        if self._mode == 'TRAIN':
+            ss_rois_num_each = int(cls_score.shape[0] / 4)
+            assert ss_rois_num_each == pi
+            det_prob1 = F.softmax(det_score[ss_rois_num_each * 0: ss_rois_num_each * 1], dim=0)
+            det_prob2 = F.softmax(det_score[ss_rois_num_each * 1: ss_rois_num_each * 2], dim=0)
+            det_prob3 = F.softmax(det_score[ss_rois_num_each * 2: ss_rois_num_each * 3], dim=0)
+            det_prob4 = F.softmax(det_score[ss_rois_num_each * 3: ss_rois_num_each * 4], dim=0)
+            det_prob = torch.cat((det_prob1, det_prob2, det_prob3, det_prob4))
+        else:
+            det_prob = F.softmax(det_score, dim=0)
+
+        det_cls_prob_product = torch.mul(cls_score, det_prob)
+        if self._mode == 'TRAIN':
+            ss_rois_num_each = int(det_cls_prob_product.shape[0] / 4)
+            assert ss_rois_num_each == pi
+            det_cls_prob1 = torch.sum(det_cls_prob_product[ss_rois_num_each * 0: ss_rois_num_each * 1], 0)
+            det_cls_prob2 = torch.sum(det_cls_prob_product[ss_rois_num_each * 1: ss_rois_num_each * 2], 0)
+            det_cls_prob3 = torch.sum(det_cls_prob_product[ss_rois_num_each * 2: ss_rois_num_each * 3], 0)
+            det_cls_prob4 = torch.sum(det_cls_prob_product[ss_rois_num_each * 3: ss_rois_num_each * 4], 0)
+            det_cls_prob = torch.stack([det_cls_prob1, det_cls_prob2, det_cls_prob3, det_cls_prob4])
+            det_cls_prob_product2 = torch.mul(cls_prob, det_prob)
+        else:
+            det_cls_prob = torch.sum(det_cls_prob_product, 0)
+            det_cls_prob_product2 = torch.mul(cls_prob, det_prob)
+
+        # bbox_pred = self.bbox_pred_net(fc7)
+        bbox_pred = torch.zeros(cls_prob.shape[0], 80)
+        self._predictions["bbox_pred"] = bbox_pred
+        self._predictions['det_cls_prob'] = det_cls_prob
+        self._predictions['det_cls_prob_product'] = det_cls_prob_product2
+
+        roi_labels_1, keep_inds_1, fg_num_1, bg_num_1 = get_refine_supervision_ac4_IA(det_cls_prob_product2,
+                                                                       self._image_gt_summaries['ss_boxes_input'][0],
+                                                                       self._image_gt_summaries['image_level_label'],
+                                                                       self._im_info)
+        roi_labels_1_new = np.vstack((roi_labels_1[0], roi_labels_1[1], roi_labels_1[2], roi_labels_1[3]))
+        keep_inds_1_new = np.concatenate((keep_inds_1[0], keep_inds_1[1] + pi, keep_inds_1[2] + pi*2, keep_inds_1[3] + pi*3))
+        fg_num_1_new = fg_num_1[0] + fg_num_1[1] + fg_num_1[2] + fg_num_1[3]
+        bg_num_1_new = bg_num_1[0] + bg_num_1[1] + bg_num_1[2] + bg_num_1[3]
+
+        roi_labels_2, keep_inds_2, fg_num_2, bg_num_2 = get_refine_supervision_ac4_IA(refine_prob_1_new,
+                                                                       self._image_gt_summaries['ss_boxes_input'][0],
+                                                                       self._image_gt_summaries['image_level_label'],
+                                                                       self._im_info)
+        roi_labels_2_new = np.vstack((roi_labels_2[0], roi_labels_2[1], roi_labels_2[2], roi_labels_2[2]))
+        keep_inds_2_new = np.concatenate((keep_inds_2[0], keep_inds_2[1] + pi, keep_inds_2[2] + pi*2, keep_inds_2[3] + pi*3))
+        fg_num_2_new = fg_num_2[0] + fg_num_2[1] + fg_num_2[2] + fg_num_2[3]
+        bg_num_2_new = bg_num_2[0] + bg_num_2[1] + bg_num_2[2] + bg_num_2[3]
+
+        self.eval()
+        gt = np.argmax(roi_labels_1_new, axis=1)
+        mask_1 = self._inverted_attention(bbox_feats_new, gt, keep_inds_1_new, 1, step, fg_num_1_new, bg_num_1_new)
+        gt = np.argmax(roi_labels_2_new, axis=1)
+        mask_2 = self._inverted_attention(bbox_feats_new, gt, keep_inds_2_new, 2, step, fg_num_2_new, bg_num_2_new)
+        self.train()
+
+        mask_1 = Variable(mask_1, requires_grad=True)
+        mask_2 = Variable(mask_2, requires_grad=True)
+        pool5_roi_1 = pool5_roi * mask_1
+        pool5_roi_2 = pool5_roi * mask_2
+        fc7_roi_1 = self._head_to_tail(pool5_roi_1)
+        fc7_roi_2 = self._head_to_tail(pool5_roi_2)
+        refine_score_1 = self.refine_net_1(fc7_roi_1)
+        refine_score_2 = self.refine_net_2(fc7_roi_2)
+
+        refine_prob_1 = F.softmax(refine_score_1, dim=1)
+        refine_prob_2 = F.softmax(refine_score_2, dim=1)
+        # refine_prob_3 = F.softmax(refine_score_3, dim=1)
+
+        self._predictions['refine_prob_1'] = refine_prob_1
+        self._predictions['refine_prob_2'] = refine_prob_2
+        # self._predictions['refine_prob_3'] = refine_prob_3
+        return roi_labels_1, keep_inds_1, roi_labels_2, keep_inds_2, bbox_pred
+
+    def _image_to_head(self):
+        raise NotImplementedError
+
+    def _head_to_tail(self, pool5):
+        raise NotImplementedError
+
+    def create_architecture(self, num_classes, tag=None, anchor_scales=(8, 16, 32), anchor_ratios=(0.5, 1, 2)):
+        self._tag = tag
+        self._num_classes = num_classes
+        self._anchor_scales = anchor_scales
+        self._num_scales = len(anchor_scales)
+        self._anchor_ratios = anchor_ratios
+        self._num_ratios = len(anchor_ratios)
+        self._num_anchors = self._num_scales * self._num_ratios
+        assert tag != None
+        # Initialize layers
+        self._init_modules()
+
+    def _init_modules(self):
+        self._init_head_tail()
+        self.cls_score_net = nn.Linear(self._fc7_channels, self._num_classes)
+        self.det_score_net = nn.Linear(self._fc7_channels, self._num_classes)
+        #self.bbox_pred_net = nn.Linear(self._fc7_channels, self._num_classes)
+        self.bbox_pred_net = nn.Linear(self._fc7_channels, self._num_classes * 4)
+        self.refine_net_1 = nn.Linear(self._fc7_channels, self._num_classes + 1)
+        self.refine_net_2 = nn.Linear(self._fc7_channels, self._num_classes + 1)
+        # self.refine_net_3 = nn.Linear(self._fc7_channels, self._num_classes + 1)
+        # self.theta = nn.Conv2d(in_channels=512, out_channels=256,
+        #                        kernel_size=1, stride=1, padding=0)
+        # self.phi = nn.Conv2d(in_channels=512, out_channels=256,
+        #                      kernel_size=1, stride=1, padding=0)
+        # self.g = nn.Conv2d(in_channels=512, out_channels=256,
+        #                      kernel_size=1, stride=1, padding=0)
+        # self.W = nn.Conv2d(in_channels=256, out_channels=512,
+        #                    kernel_size=1, stride=1, padding=0)
+        # nn.init.constant_(self.W.weight, 0)
+        # nn.init.constant_(self.W.bias, 0)
+        self.init_weights()
+
+    def _run_summary_op(self, val=False):
+        """
+        Run the summary operator: feed the placeholders with corresponding newtork outputs(activations)
+        """
+        summaries = []
+        # Add image gt
+        summaries.append(self._add_gt_image_summary())
+        # Add event_summaries
+        for key, var in self._event_summaries.items():  # __event_summaries is equal to loss itmes
+            summaries.append(tb.summary.scalar(key, var.item()))
+        self._event_summaries = {}
+        return summaries
+
+    def _predict_train(self, ss_boxes, step):
+        torch.backends.cudnn.benchmark = False
+        net_conv = self._image_to_head()
+        i = 0
+        rois = torch.from_numpy(ss_boxes[i]).to(self._device)
+        pool5_roi_0 = self.RoIRingPool(net_conv[i][0:1, :], rois[0, :])
+        pool5_context_0 = self.RoIRingPool_context(net_conv[i][0:1, :], rois[0, :])
+        pool5_frame_0 = self.RoIRingPool_frame(net_conv[i][0:1, :], rois[0, :])
+        pool5_roi_flip_0 = self.RoIRingPool(net_conv[i][1:2, :], rois[1, :])
+        pool5_context_flip_0 = self.RoIRingPool_context(net_conv[i][1:2, :], rois[1, :])
+        pool5_frame_flip_0 = self.RoIRingPool_frame(net_conv[i][1:2, :], rois[1, :])
+        i = 1
+        rois = torch.from_numpy(ss_boxes[i]).to(self._device)
+        pool5_roi_1 = self.RoIRingPool(net_conv[i][0:1, :], rois[0, :])
+        pool5_context_1 = self.RoIRingPool_context(net_conv[i][0:1, :], rois[0, :])
+        pool5_frame_1 = self.RoIRingPool_frame(net_conv[i][0:1, :], rois[0, :])
+        pool5_roi_flip_1 = self.RoIRingPool(net_conv[i][1:2, :], rois[1, :])
+        pool5_context_flip_1 = self.RoIRingPool_context(net_conv[i][1:2, :], rois[1, :])
+        pool5_frame_flip_1 = self.RoIRingPool_frame(net_conv[i][1:2, :], rois[1, :])
+
+        pool5_roi = torch.cat((pool5_roi_0, pool5_roi_flip_0, pool5_roi_1, pool5_roi_flip_1))
+        pool5_context = torch.cat((pool5_context_0, pool5_context_flip_0, pool5_context_1, pool5_context_flip_1))
+        pool5_frame = torch.cat((pool5_frame_0, pool5_frame_flip_0, pool5_frame_1, pool5_frame_flip_1))
+
+        if self._mode == 'TRAIN':
+            torch.backends.cudnn.benchmark = True  # benchmark because now the input size are fixed
+        fc7_roi = self._head_to_tail(pool5_roi)
+        fc7_context = self._head_to_tail(pool5_context)
+        fc7_frame = self._head_to_tail(pool5_frame)
+        if self.ca_iw:
+            rois = pool5_roi
+        else:
+            rois = None
+
+        roi_labels_1, keep_inds_1, \
+        roi_labels_2, keep_inds_2, bbox_pred = self._region_classification_train(pool5_roi, fc7_roi,fc7_context, fc7_frame, step)
+        return roi_labels_1, keep_inds_1, roi_labels_2, keep_inds_2, bbox_pred, rois
+
+    def _predict_test(self, ss_boxes):
+        torch.backends.cudnn.benchmark = False
+        net_conv = self._image_to_head()
+        ss_rois = torch.from_numpy(ss_boxes).to(self._device)
+        rois = ss_rois
+        self._predictions["rois"] = rois
+        pool5_roi = self.RoIRingPool(net_conv, rois)
+        pool5_context = self.RoIRingPool_context(net_conv, rois)
+        pool5_frame = self.RoIRingPool_frame(net_conv, rois)
+
+        if self._mode == 'TRAIN':
+            torch.backends.cudnn.benchmark = True  # benchmark because now the input size are fixed
+        fc7_roi = self._head_to_tail(pool5_roi)
+        fc7_context = self._head_to_tail(pool5_context)
+        fc7_frame = self._head_to_tail(pool5_frame)
+
+        cls_prob, det_prob, bbox_pred, cls_det_prob_product, det_cls_prob = self._region_classification_test(fc7_roi, fc7_context, fc7_frame)
+        return rois, cls_prob, det_prob, bbox_pred, cls_det_prob_product, det_cls_prob
+
+    def forward(self, image, image_level_label, im_info, gt_boxes=None, ss_boxes=None, step=None, mode='TRAIN'):
+        self._image_gt_summaries['image'] = image
+        self._image_gt_summaries['image_level_label'] = image_level_label
+        self._image_gt_summaries['gt_boxes'] = gt_boxes
+        self._image_gt_summaries['im_info'] = im_info
+        self._mode = mode
+        self._im_info = im_info
+        self._image_level_label = torch.from_numpy(image_level_label) if image_level_label is not None else None
+
+        if mode == 'TEST':
+            self._image_gt_summaries['ss_boxes'] = ss_boxes
+            self._image = torch.from_numpy(image.transpose([0, 3, 1, 2]).copy()).to(self._device)
+            self._gt_boxes = torch.from_numpy(gt_boxes).to(self._device) if gt_boxes is not None else None
+            self.ss_boxes_indexes = self.return_ss_boxes(np.arange(ss_boxes.shape[0]), mode)
+            rois, cls_prob, det_prob, bbox_pred, cls_det_prob_product, det_cls_prob = self._predict_test(ss_boxes[self.ss_boxes_indexes, :])
+            bbox_pred = bbox_pred[:, :80]
+            stds = bbox_pred.data.new(cfg.TRAIN.BBOX_NORMALIZE_STDS).repeat(self._num_classes).unsqueeze(0).expand_as(bbox_pred)
+            means = bbox_pred.data.new(cfg.TRAIN.BBOX_NORMALIZE_MEANS).repeat(self._num_classes).unsqueeze(0).expand_as(bbox_pred)
+            self._predictions["bbox_pred"] = bbox_pred.mul(stds).add(means)
+        else:
+            ss_boxes_all = []
+            self._image = []
+            self._image_gt_summaries['ss_boxes_input'] = []
+            self._image_level_label = torch.from_numpy(image_level_label) if image_level_label is not None else None
+            self.ss_boxes_indexes = self.return_ss_boxes(np.arange(ss_boxes[0].shape[0]), mode)
+            for i in range(2):
+                image_org = torch.from_numpy(image[i].transpose([0, 3, 1, 2]).copy()).to(self._device)
+                self._image.append(image_org)
+                ss_boxes_input = np.stack((ss_boxes[i * 2], ss_boxes[i * 2 + 1]))
+                ss_boxes_all.append(ss_boxes_input[:, self.ss_boxes_indexes, :])
+            self._image_gt_summaries['ss_boxes_input'] = ss_boxes_all
+
+            roi_labels_1, keep_inds_1, roi_labels_2, keep_inds_2, bbox_pred, rois = self._predict_train(ss_boxes_all, step)
+            bbox_pred = bbox_pred[:, :80]
+            self._add_losses(roi_labels_1, keep_inds_1, roi_labels_2, keep_inds_2, step=step, rois=rois)
+
+    def return_ss_boxes(self, boxes_index, mode='TRAIN'):
+        if mode == 'TEST':
+            return boxes_index
+        box_num = min(1000, len(boxes_index))
+        indexes = np.random.choice(boxes_index, size=box_num, replace=False)
+        return indexes
+
+    def init_weights(self):
+        def normal_init(m, mean, stddev, truncated=False):
+            """
+            weight initalizer: truncated normal and random normal.
+            """
+            if truncated:
+                m.weight.data.normal_().fmod_(2).mul_(stddev).add_(mean)  # not a perfect approximation
+            else:
+                m.weight.data.normal_(mean, stddev)
+            m.bias.data.zero_()
+
+        # normal_init(self.rpn_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init(self.rpn_cls_score_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init(self.rpn_bbox_pred_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        normal_init(self.cls_score_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        normal_init(self.det_score_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        normal_init(self.bbox_pred_net, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init(self.bbox_pred_net, 0, 0.001, cfg.TRAIN.TRUNCATED)
+        normal_init(self.refine_net_1, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        normal_init(self.refine_net_2, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init(self.refine_net_3, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init1(self.theta, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init1(self.phi, 0, 0.01, cfg.TRAIN.TRUNCATED)
+        # normal_init1(self.g, 0, 0.01, cfg.TRAIN.TRUNCATED)
+
+    # Extract the head feature maps, for example for vgg16 it is conv5_3
+    # only useful during testing mode
+    def extract_head(self, image):
+        feat = self._layers["head"](torch.from_numpy(image.transpose([0, 3, 1, 2])).to(self._device))
+        return feat
+
+    # only useful during testing mode
+    def test_image(self, image, im_info, ss_boxes):
+        self.eval()
+        with torch.no_grad():
+            self.forward(image, None, im_info, None, ss_boxes, mode='TEST')
+
+        bbox_pred, rois, det_cls_prob, det_cls_prob_product, refine_prob_1, refine_prob_2 = \
+            self._predictions['bbox_pred'].data.cpu().numpy(), \
+            self._predictions['rois'].data.cpu().numpy(), \
+            self._predictions['det_cls_prob'].data.cpu().numpy(), \
+            self._predictions['det_cls_prob_product'].data.cpu().numpy(), \
+            self._predictions['refine_prob_1'].data.cpu().numpy(), \
+            self._predictions['refine_prob_2'].data.cpu().numpy()
+
+        return bbox_pred, rois, det_cls_prob, det_cls_prob_product, refine_prob_1[:, 1:], refine_prob_2[:, 1:]
+
+    def delete_intermediate_states(self):
+        # Delete intermediate result to save memory
+        for d in [self._losses, self._predictions]:
+            for k in list(d):
+                del d[k]
+
+    def get_summary(self, blobs, step=None):
+        self.eval()
+        self.forward(blobs['data'], blobs['image_level_labels'], blobs['im_info'], blobs['gt_boxes'], blobs['ss_boxes'], step)
+        self.train()
+        summary = self._run_summary_op(True)
+        return summary
+
+    def train_step(self, blobs, train_op, step):
+        self.forward(blobs['data'], blobs['image_level_labels'], blobs['im_info'], blobs['gt_boxes'], blobs['ss_boxes'], step)
+        cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, loss = self._losses['cls_det_loss'].item(), \
+                                                                             self._losses['refine_loss_1'].item(), \
+                                                                             self._losses['refine_loss_2'].item(), \
+                                                                             self._losses['consistency_loss'].item(), \
+                                                                             self._losses['total_loss'].item()
+        train_op.zero_grad()
+        self._losses['total_loss'].backward()
+        train_op.step()
+
+        self.delete_intermediate_states()
+        #torch.cuda.empty_cache()
+
+        return cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, loss
+
+    def train_step_with_summary(self, blobs, train_op, step):
+        self.forward(blobs['data'], blobs['image_level_labels'], blobs['im_info'], blobs['gt_boxes'], blobs['ss_boxes'], step)
+        cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, loss = self._losses["cls_det_loss"].item(), \
+                                                           self._losses['refine_loss_1'].item(), \
+                                                           self._losses['refine_loss_2'].item(), \
+                                                           self._losses['consistency_loss'].item(), \
+                                                           self._losses['total_loss'].item()
+        train_op.zero_grad()
+        self._losses['total_loss'].backward()
+        train_op.step()
+        # summary = self._run_summary_op()
+        summary = 0
+        self.delete_intermediate_states()
+        return cls_det_loss, refine_loss_1, refine_loss_2, consistency_loss, loss, summary
+
+
+    def train_step_no_return(self, blobs, train_op):
+        self.forward(blobs['data'], blobs['im_info'], blobs['gt_boxes'], blobs['ss_boxes'])
+        train_op.zero_grad()
+        self._losses['total_loss'].backward()
+        train_op.step()
+        self.delete_intermediate_states()
+
+    def load_state_dict(self, state_dict):
+        """
+        Because we remove the definition of fc layer in resnet now, it will fail when loading
+        the model trained before.
+        To provide back compatibility, we overwrite the load_state_dict
+        """
+        nn.Module.load_state_dict(self, {k: state_dict[k] for k in list(self.state_dict())})
+
+# ----------------------------------------------------------------------------------------------------------------------
+def _get_top_ranking_propoals(probs):
+    """Get top ranking proposals by k-means"""
+    kmeans = KMeans(n_clusters=3, random_state=3).fit(probs)
+    high_score_label = np.argmax(kmeans.cluster_centers_)
+    index = np.where(kmeans.labels_ == high_score_label)[0]
+    if len(index) == 0:
+        index = np.array([np.argmax(probs)])
+    return index
+
+
+def _build_graph(boxes, iou_threshold):
+    """Build graph based on box IoU"""
+    overlaps = bbox_overlaps(
+        boxes.astype(dtype=np.float32, copy=False),
+        boxes.astype(dtype=np.float32, copy=False))
+    return (overlaps > iou_threshold).astype(np.float32)
+
+
+def _get_graph_centers(boxes, cls_prob, im_labels):
+    """Get graph centers."""
+    num_images, num_classes = im_labels.shape
+    assert num_images == 1, 'batch size shoud be equal to 1'
+    im_labels_tmp = im_labels[0, :].copy()
+    gt_boxes = np.zeros((0, 4), dtype=np.float32)
+    gt_classes = np.zeros((0, 1), dtype=np.int32)
+    gt_scores = np.zeros((0, 1), dtype=np.float32)
+    for i in xrange(num_classes):
+        if im_labels_tmp[i] == 1:
+            cls_prob_tmp = cls_prob[:, i].copy()
+            idxs = np.where(cls_prob_tmp >= 0)[0]
+            if idxs.shape[0] == 0:
+                print('kmeans problem')
+                continue
+            idxs_tmp = _get_top_ranking_propoals(cls_prob_tmp[idxs].reshape(-1, 1))
+            idxs = idxs[idxs_tmp]
+            boxes_tmp = boxes[idxs, :].copy()
+            cls_prob_tmp = cls_prob_tmp[idxs]
+            graph = _build_graph(boxes_tmp, 0.4)
+
+            keep_idxs = []
+            gt_scores_tmp = []
+            count = cls_prob_tmp.size
+            while True:
+                order = np.sum(graph, axis=1).argsort()[::-1]
+                tmp = order[0]
+                keep_idxs.append(tmp)
+                inds = np.where(graph[tmp, :] > 0)[0]
+                gt_scores_tmp.append(np.max(cls_prob_tmp[inds]))
+
+                graph[:, inds] = 0
+                graph[inds, :] = 0
+                count = count - len(inds)
+                if count <= 5:
+                    break
+
+            gt_boxes_tmp = boxes_tmp[keep_idxs, :].copy()
+            gt_scores_tmp = np.array(gt_scores_tmp).copy()
+
+            keep_idxs_new = np.argsort(gt_scores_tmp)[-1:(-1 - min(len(gt_scores_tmp), 5)):-1]
+
+            gt_boxes = np.vstack((gt_boxes, gt_boxes_tmp[keep_idxs_new, :]))
+            gt_scores = np.vstack((gt_scores,gt_scores_tmp[keep_idxs_new].reshape(-1, 1)))
+            gt_classes = np.vstack((gt_classes,(i + 1) * np.ones((len(keep_idxs_new), 1), dtype=np.int32)))
+
+            # If a proposal is chosen as a cluster center,
+            # we simply delete a proposal from the candidata proposal pool,
+            # because we found that the results of different strategies are similar and this strategy is more efficient
+            cls_prob = np.delete(cls_prob.copy(), idxs[keep_idxs][keep_idxs_new], axis=0)
+            boxes = np.delete(boxes.copy(), idxs[keep_idxs][keep_idxs_new], axis=0)
+
+    proposals = {'gt_boxes': gt_boxes,
+                 'gt_classes': gt_classes,
+                 'gt_scores': gt_scores}
+
+    return proposals
+
+
+def _get_proposal_clusters(all_rois, proposals, im_labels):
+    """Generate a random sample of RoIs comprising foreground and background
+    examples.
+    """
+    num_images, num_classes = im_labels.shape
+    assert num_images == 1, 'batch size shoud be equal to 1'
+    # overlaps: (rois x gt_boxes)
+    gt_boxes = proposals['gt_boxes']
+    gt_labels = proposals['gt_classes']
+    #gt_scores = proposals['gt_scores']
+    overlaps = bbox_overlaps(
+        all_rois.astype(dtype=np.float32, copy=False),
+        gt_boxes.astype(dtype=np.float32, copy=False))
+    gt_assignment = overlaps.argmax(axis=1)
+    max_overlaps = overlaps.max(axis=1)
+    labels = gt_labels[gt_assignment, 0]
+    # cls_loss_weights = gt_scores[gt_assignment, 0]
+
+    # # Select foreground RoIs as those with >= FG_THRESH overlap
+    # fg_inds = np.where(max_overlaps >= cfg.TRAIN.FG_THRESH)[0]
+    #
+    # # Select background RoIs as those with < FG_THRESH overlap
+    # bg_inds = np.where(max_overlaps < cfg.TRAIN.FG_THRESH)[0]
+    #
+    # ig_inds = np.where(max_overlaps < cfg.TRAIN.BG_THRESH)[0]
+
+    # cls_loss_weights[ig_inds] = 0.0
+    #
+    # labels[bg_inds] = 0
+    # gt_assignment[bg_inds] = -1
+    #
+    # img_cls_loss_weights = np.zeros(gt_boxes.shape[0], dtype=np.float32)
+    # pc_probs = np.zeros(gt_boxes.shape[0], dtype=np.float32)
+    # pc_labels = np.zeros(gt_boxes.shape[0], dtype=np.int32)
+    # pc_count = np.zeros(gt_boxes.shape[0], dtype=np.int32)
+    #
+    # for i in xrange(gt_boxes.shape[0]):
+    #     po_index = np.where(gt_assignment == i)[0]
+    #     img_cls_loss_weights[i] = np.sum(cls_loss_weights[po_index])
+    #     pc_labels[i] = gt_labels[i, 0]
+    #     pc_count[i] = len(po_index)
+    #     pc_probs[i] = np.average(cls_prob[po_index, pc_labels[i]])
+    return max_overlaps, labels
+
+
+def get_refine_supervision_ac4_IA(refine_prob, ss_boxes, image_level_label, im_info=None):
+    '''
+    refine_prob: num_box x 20 or num_box x 21
+    ss_boxes; num_box x 4
+    image_level_label: 1 dim vector with 20 elements
+    '''
+    keep_inds_list = []
+    roi_labels_list = []
+    pi = ss_boxes.shape[1]
+    fg_len = []
+    bg_len = []
+
+    refine_prob_each = (refine_prob[0:pi, :] + refine_prob[pi:pi*2, :] + refine_prob[pi*2:pi*3, :] + refine_prob[pi*3:pi*4, :]) / 4.0
+    for i in range(1):
+        ss_boxes_each = ss_boxes[i, :]
+
+        cls_prob = refine_prob_each.data.cpu().numpy()
+        boxes = ss_boxes_each[:, 1:].copy()
+
+        if refine_prob.shape[1] == image_level_label.shape[1] + 1:
+            cls_prob = cls_prob[:, 1:]
+        roi_labels = np.zeros([pi, image_level_label.shape[1] + 1], dtype=np.int32)
+        roi_labels[:, 0] = 1  # the 0th elements is the bg
+        roi_weights = np.zeros((pi, 1), dtype=np.float32)  # num_box x 1 weights of the rois
+
+        eps = 1e-9
+        cls_prob[cls_prob < eps] = eps
+        cls_prob[cls_prob > 1 - eps] = 1 - eps
+        proposals = _get_graph_centers(boxes.copy(), cls_prob.copy(), image_level_label.copy())
+        # proposals_list.append(proposals)
+
+        max_overlaps, labels = _get_proposal_clusters(boxes.copy(), proposals, image_level_label.copy())
+
+        fg_inds = np.where(max_overlaps > cfg.TRAIN.MIL_FG_THRESH)[0]
+
+        roi_labels[fg_inds, labels[fg_inds]] = 1
+        roi_labels[fg_inds, 0] = 0
+
+        bg_inds = (np.array(max_overlaps >= cfg.TRAIN.MIL_BG_THRESH_LO, dtype=np.int32) + np.array(
+            max_overlaps < cfg.TRAIN.MIL_BG_THRESH_HI, dtype=np.int32) == 2).nonzero()[0]
+
+        for m in range(4):
+            if len(fg_inds) > 0 and len(bg_inds) > 0:
+                fg_rois_num = min(cfg.TRAIN.MIL_NUM_FG, len(fg_inds))
+                fg_inds_tmp = fg_inds[np.random.choice(np.arange(0, len(fg_inds)), size=int(fg_rois_num), replace=False)]
+                bg_rois_num = min(cfg.TRAIN.MIL_NUM_BG, len(bg_inds))
+                bg_inds_tmp = bg_inds[np.random.choice(np.arange(0, len(bg_inds)), size=int(bg_rois_num), replace=False)]
+            elif len(fg_inds) > 0:
+                fg_rois_num = min(cfg.TRAIN.MIL_NUM_FG, len(fg_inds))
+                fg_inds_tmp = fg_inds[np.random.choice(np.arange(0, len(fg_inds)), size=int(fg_rois_num), replace=False)]
+                bg_inds_tmp = bg_inds
+            elif len(bg_inds) > 0:
+                bg_rois_num = min(cfg.TRAIN.MIL_NUM_BG, len(bg_inds))
+                bg_inds_tmp = bg_inds[np.random.choice(np.arange(0, len(bg_inds)), size=int(bg_rois_num), replace=False)]
+                fg_inds_tmp = fg_inds
+            else:
+                import pdb
+                pdb.set_trace()
+
+            for n in range(1):
+                keep_inds = np.concatenate([fg_inds_tmp, bg_inds_tmp])
+                keep_inds_list.append(keep_inds)
+                #roi_labels_list.append(roi_labels[keep_inds, :])
+                roi_labels_list.append(roi_labels)
+                fg_len.append(len(fg_inds_tmp))
+                bg_len.append(len(bg_inds_tmp))
+
+    return roi_labels_list, keep_inds_list, fg_len, bg_len
+
diff --git a/lib/nets/resnet_v1.py b/lib/nets/resnet_v1.py
new file mode 100644
index 0000000..d051bb7
--- /dev/null
+++ b/lib/nets/resnet_v1.py
@@ -0,0 +1,186 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Zheqi He and Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from nets.network import Network
+from model.config import cfg
+
+import utils.timer
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+import math
+import torch.utils.model_zoo as model_zoo
+
+import torchvision
+from torchvision.models.resnet import BasicBlock, Bottleneck
+
+class ResNet(torchvision.models.resnet.ResNet):
+  def __init__(self, block, layers, num_classes=1000):
+    self.inplanes = 64
+    super(ResNet, self).__init__(block, layers, num_classes)
+    # change to match the caffe resnet
+    for i in range(2, 4):
+      getattr(self, 'layer%d'%i)[0].conv1.stride = (2,2)
+      getattr(self, 'layer%d'%i)[0].conv2.stride = (1,1)
+    # use stride 1 for the last conv4 layer (same as tf-faster-rcnn)
+    self.layer4[0].conv2.stride = (1,1)
+    self.layer4[0].downsample[0].stride = (1,1)
+
+    del self.avgpool, self.fc
+
+
+def resnet18(pretrained=False):
+  """Constructs a ResNet-18 model.
+  Args:
+    pretrained (bool): If True, returns a model pre-trained on ImageNet
+  """
+  model = ResNet(BasicBlock, [2, 2, 2, 2])
+  if pretrained:
+    model.load_state_dict(model_zoo.load_url(model_urls['resnet18']))
+  return model
+
+
+def resnet34(pretrained=False):
+  """Constructs a ResNet-34 model.
+  Args:
+    pretrained (bool): If True, returns a model pre-trained on ImageNet
+  """
+  model = ResNet(BasicBlock, [3, 4, 6, 3])
+  if pretrained:
+    model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))
+  return model
+
+
+def resnet50(pretrained=False):
+  """Constructs a ResNet-50 model.
+  Args:
+    pretrained (bool): If True, returns a model pre-trained on ImageNet
+  """
+  model = ResNet(Bottleneck, [3, 4, 6, 3])
+  if pretrained:
+    model.load_state_dict(model_zoo.load_url(model_urls['resnet50']))
+  return model
+
+
+def resnet101(pretrained=False):
+  """Constructs a ResNet-101 model.
+  Args:
+    pretrained (bool): If True, returns a model pre-trained on ImageNet
+  """
+  model = ResNet(Bottleneck, [3, 4, 23, 3])
+  if pretrained:
+    model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
+  return model
+
+
+def resnet152(pretrained=False):
+  """Constructs a ResNet-152 model.
+  Args:
+    pretrained (bool): If True, returns a model pre-trained on ImageNet
+  """
+  model = ResNet(Bottleneck, [3, 8, 36, 3])
+  if pretrained:
+    model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
+  return model
+
+class resnetv1(Network):
+  def __init__(self, num_layers=50):
+    Network.__init__(self)
+    self._feat_stride = [16, ]
+    self._feat_compress = [1. / float(self._feat_stride[0]), ]
+    self._num_layers = num_layers
+    self._net_conv_channels = 1024
+    self._fc7_channels = 2048
+
+  def _crop_pool_layer(self, bottom, rois):
+    return Network._crop_pool_layer(self, bottom, rois, cfg.RESNET.MAX_POOL)
+
+  def _image_to_head(self):
+    net_conv = self._layers['head'](self._image)
+    self._act_summaries['conv'] = net_conv
+
+    return net_conv
+
+  def _head_to_tail(self, pool5):
+    fc7 = self.resnet.layer4(pool5).mean(3).mean(2) # average pooling after layer4
+    return fc7
+
+  def _init_head_tail(self):
+    # choose different blocks for different number of layers
+    if self._num_layers == 50:
+      self.resnet = resnet50()
+
+    elif self._num_layers == 101:
+      self.resnet = resnet101()
+
+    elif self._num_layers == 152:
+      self.resnet = resnet152()
+
+    else:
+      # other numbers are not supported
+      raise NotImplementedError
+
+    # Fix blocks 
+    for p in self.resnet.bn1.parameters(): p.requires_grad=False
+    for p in self.resnet.conv1.parameters(): p.requires_grad=False
+    assert (0 <= cfg.RESNET.FIXED_BLOCKS < 4)
+    if cfg.RESNET.FIXED_BLOCKS >= 3:
+      for p in self.resnet.layer3.parameters(): p.requires_grad=False
+    if cfg.RESNET.FIXED_BLOCKS >= 2:
+      for p in self.resnet.layer2.parameters(): p.requires_grad=False
+    if cfg.RESNET.FIXED_BLOCKS >= 1:
+      for p in self.resnet.layer1.parameters(): p.requires_grad=False
+
+    def set_bn_fix(m):
+      classname = m.__class__.__name__
+      if classname.find('BatchNorm') != -1:
+        for p in m.parameters(): p.requires_grad=False
+
+    self.resnet.apply(set_bn_fix)
+
+    # Build resnet.
+    self._layers['head'] = nn.Sequential(self.resnet.conv1, self.resnet.bn1,self.resnet.relu, 
+      self.resnet.maxpool,self.resnet.layer1,self.resnet.layer2,self.resnet.layer3)
+
+  def train(self, mode=True):
+    # Override train so that the training mode is set as we want
+    nn.Module.train(self, mode)
+    if mode:
+      # Set fixed blocks to be in eval mode (not really doing anything)
+      self.resnet.eval()
+      if cfg.RESNET.FIXED_BLOCKS <= 3:
+        self.resnet.layer4.train()
+      if cfg.RESNET.FIXED_BLOCKS <= 2:
+        self.resnet.layer3.train()
+      if cfg.RESNET.FIXED_BLOCKS <= 1:
+        self.resnet.layer2.train()
+      if cfg.RESNET.FIXED_BLOCKS == 0:
+        self.resnet.layer1.train()
+
+      # Set batchnorm always in eval mode during training
+      def set_bn_eval(m):
+        classname = m.__class__.__name__
+        if classname.find('BatchNorm') != -1:
+          m.eval()
+
+      self.resnet.apply(set_bn_eval)
+
+  def load_pretrained_cnn(self, state_dict):
+    self.resnet.load_state_dict({k: state_dict[k] for k in list(self.resnet.state_dict())})
+
+
+
+
+
+
+
+
+
diff --git a/lib/nets/vgg16.py b/lib/nets/vgg16.py
new file mode 100644
index 0000000..55249d4
--- /dev/null
+++ b/lib/nets/vgg16.py
@@ -0,0 +1,67 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from nets.network import Network
+from model.config import cfg
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+import math
+import torchvision.models as models
+
+class MELM_vgg16(Network):
+  def __init__(self):
+    Network.__init__(self)
+    self._feat_stride = [16, ]
+    self._feat_compress = [1. / float(self._feat_stride[0]), ]
+    self._net_conv_channels = 512
+    self._fc7_channels = 4096
+
+  def _init_head_tail(self):
+    self.vgg = models.vgg16()
+    # Remove fc8
+    self.vgg.classifier = nn.Sequential(*list(self.vgg.classifier._modules.values())[:-1])
+
+    # Fix the layers before conv3:
+    for layer in range(10):
+      for p in self.vgg.features[layer].parameters(): p.requires_grad = False
+
+    # not using the last maxpool layer
+    self._layers['head'] = nn.Sequential(*list(self.vgg.features._modules.values())[:-1])
+
+    # ------- parallel Gpu-----------
+    self._layers['head'] = torch.nn.DataParallel(self._layers['head'])
+
+  def _image_to_headtest(self):
+    net_conv = self._layers['head'](self._image)
+    return net_conv
+
+  def _image_to_head(self):
+    if self.training:
+      net_conv = []
+      for i in range(2):
+        net_conv.append(self._layers['head'](self._image[i]))
+    else:
+      net_conv = self._layers['head'](self._image)
+    return net_conv
+
+  def _head_to_tail(self, pool5):
+    pool5_flat = pool5.view(pool5.size(0), -1)
+    fc7 = self.vgg.classifier(pool5_flat)
+    self._predictions['fc7'] = fc7
+
+    return fc7
+
+  def load_pretrained_cnn(self, state_dict):
+    self.vgg.load_state_dict({k:v for k,v in state_dict.items() if k in self.vgg.state_dict()})
+    #self.vgg.classifier = torch.nn.DataParallel(self.vgg.classifier)
+
+
diff --git a/lib/ops/__init__.py b/lib/ops/__init__.py
new file mode 100644
index 0000000..f319180
--- /dev/null
+++ b/lib/ops/__init__.py
@@ -0,0 +1,15 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 29 15:19:42 2019
+
+@author: vasgaoweithu
+"""
+
+from .nms import nms, soft_nms
+from .roi_align import RoIAlign
+from .roi_pool import RoIPool
+from .roi_crop import RoICrop
+from .roi_ring_pool import RoIRingPool
+
+__all__ = ['nms', 'soft_nms', 'RoIAlign', 'RoIPool', 'RoICrop', 'RoIRingPool']
\ No newline at end of file
diff --git a/lib/ops/__pycache__/__init__.cpython-37.pyc b/lib/ops/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..50b9d83
Binary files /dev/null and b/lib/ops/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/nms/__init__.py b/lib/ops/nms/__init__.py
new file mode 100644
index 0000000..c440704
--- /dev/null
+++ b/lib/ops/nms/__init__.py
@@ -0,0 +1,3 @@
+from .nms_wrapper import nms, soft_nms
+
+__all__ = ['nms', 'soft_nms']
diff --git a/lib/ops/nms/__pycache__/__init__.cpython-37.pyc b/lib/ops/nms/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..ba7698b
Binary files /dev/null and b/lib/ops/nms/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/nms/__pycache__/nms_wrapper.cpython-37.pyc b/lib/ops/nms/__pycache__/nms_wrapper.cpython-37.pyc
new file mode 100644
index 0000000..f066d87
Binary files /dev/null and b/lib/ops/nms/__pycache__/nms_wrapper.cpython-37.pyc differ
diff --git a/lib/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..4862476
Binary files /dev/null and b/lib/ops/nms/nms_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..4dacae6
Binary files /dev/null and b/lib/ops/nms/nms_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/nms/nms_wrapper.py b/lib/ops/nms/nms_wrapper.py
new file mode 100644
index 0000000..03215f1
--- /dev/null
+++ b/lib/ops/nms/nms_wrapper.py
@@ -0,0 +1,78 @@
+import numpy as np
+import torch
+
+from . import nms_cpu, nms_cuda
+from .soft_nms_cpu import soft_nms_cpu
+
+
+def nms(dets, iou_thr, device_id=None):
+    """Dispatch to either CPU or GPU NMS implementations.
+
+    The input can be either a torch tensor or numpy array. GPU NMS will be used
+    if the input is a gpu tensor or device_id is specified, otherwise CPU NMS
+    will be used. The returned type will always be the same as inputs.
+
+    Arguments:
+        dets (torch.Tensor or np.ndarray): bboxes with scores.
+        iou_thr (float): IoU threshold for NMS.
+        device_id (int, optional): when `dets` is a numpy array, if `device_id`
+            is None, then cpu nms is used, otherwise gpu_nms will be used.
+
+    Returns:
+        tuple: kept bboxes and indice, which is always the same data type as
+            the input.
+    """
+    # convert dets (tensor or numpy array) to tensor
+    if isinstance(dets, torch.Tensor):
+        is_numpy = False
+        dets_th = dets
+    elif isinstance(dets, np.ndarray):
+        is_numpy = True
+        device = 'cpu' if device_id is None else 'cuda:{}'.format(device_id)
+        dets_th = torch.from_numpy(dets).to(device)
+    else:
+        raise TypeError(
+            'dets must be either a Tensor or numpy array, but got {}'.format(
+                type(dets)))
+
+    # execute cpu or cuda nms
+    if dets_th.shape[0] == 0:
+        inds = dets_th.new_zeros(0, dtype=torch.long)
+    else:
+        if dets_th.is_cuda:
+            inds = nms_cuda.nms(dets_th, iou_thr)
+        else:
+            inds = nms_cpu.nms(dets_th, iou_thr)
+
+    if is_numpy:
+        inds = inds.cpu().numpy()
+    return dets[inds, :], inds
+
+
+def soft_nms(dets, iou_thr, method='linear', sigma=0.5, min_score=1e-3):
+    if isinstance(dets, torch.Tensor):
+        is_tensor = True
+        dets_np = dets.detach().cpu().numpy()
+    elif isinstance(dets, np.ndarray):
+        is_tensor = False
+        dets_np = dets
+    else:
+        raise TypeError(
+            'dets must be either a Tensor or numpy array, but got {}'.format(
+                type(dets)))
+
+    method_codes = {'linear': 1, 'gaussian': 2}
+    if method not in method_codes:
+        raise ValueError('Invalid method for SoftNMS: {}'.format(method))
+    new_dets, inds = soft_nms_cpu(
+        dets_np,
+        iou_thr,
+        method=method_codes[method],
+        sigma=sigma,
+        min_score=min_score)
+
+    if is_tensor:
+        return dets.new_tensor(new_dets), dets.new_tensor(
+            inds, dtype=torch.long)
+    else:
+        return new_dets.astype(np.float32), inds.astype(np.int64)
diff --git a/lib/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..04e9bb9
Binary files /dev/null and b/lib/ops/nms/soft_nms_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/nms/src/nms_cpu.cpp b/lib/ops/nms/src/nms_cpu.cpp
new file mode 100644
index 0000000..f7cffb4
--- /dev/null
+++ b/lib/ops/nms/src/nms_cpu.cpp
@@ -0,0 +1,71 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+#include <torch/extension.h>
+
+template <typename scalar_t>
+at::Tensor nms_cpu_kernel(const at::Tensor& dets, const float threshold) {
+  AT_ASSERTM(!dets.type().is_cuda(), "dets must be a CPU tensor");
+
+  if (dets.numel() == 0) {
+    return at::empty({0}, dets.options().dtype(at::kLong).device(at::kCPU));
+  }
+
+  auto x1_t = dets.select(1, 0).contiguous();
+  auto y1_t = dets.select(1, 1).contiguous();
+  auto x2_t = dets.select(1, 2).contiguous();
+  auto y2_t = dets.select(1, 3).contiguous();
+  auto scores = dets.select(1, 4).contiguous();
+
+  at::Tensor areas_t = (x2_t - x1_t + 1) * (y2_t - y1_t + 1);
+
+  auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
+
+  auto ndets = dets.size(0);
+  at::Tensor suppressed_t =
+      at::zeros({ndets}, dets.options().dtype(at::kByte).device(at::kCPU));
+
+  auto suppressed = suppressed_t.data<uint8_t>();
+  auto order = order_t.data<int64_t>();
+  auto x1 = x1_t.data<scalar_t>();
+  auto y1 = y1_t.data<scalar_t>();
+  auto x2 = x2_t.data<scalar_t>();
+  auto y2 = y2_t.data<scalar_t>();
+  auto areas = areas_t.data<scalar_t>();
+
+  for (int64_t _i = 0; _i < ndets; _i++) {
+    auto i = order[_i];
+    if (suppressed[i] == 1) continue;
+    auto ix1 = x1[i];
+    auto iy1 = y1[i];
+    auto ix2 = x2[i];
+    auto iy2 = y2[i];
+    auto iarea = areas[i];
+
+    for (int64_t _j = _i + 1; _j < ndets; _j++) {
+      auto j = order[_j];
+      if (suppressed[j] == 1) continue;
+      auto xx1 = std::max(ix1, x1[j]);
+      auto yy1 = std::max(iy1, y1[j]);
+      auto xx2 = std::min(ix2, x2[j]);
+      auto yy2 = std::min(iy2, y2[j]);
+
+      auto w = std::max(static_cast<scalar_t>(0), xx2 - xx1 + 1);
+      auto h = std::max(static_cast<scalar_t>(0), yy2 - yy1 + 1);
+      auto inter = w * h;
+      auto ovr = inter / (iarea + areas[j] - inter);
+      if (ovr >= threshold) suppressed[j] = 1;
+    }
+  }
+  return at::nonzero(suppressed_t == 0).squeeze(1);
+}
+
+at::Tensor nms(const at::Tensor& dets, const float threshold) {
+  at::Tensor result;
+  AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms", [&] {
+    result = nms_cpu_kernel<scalar_t>(dets, threshold);
+  });
+  return result;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("nms", &nms, "non-maximum suppression");
+}
\ No newline at end of file
diff --git a/lib/ops/nms/src/nms_cuda.cpp b/lib/ops/nms/src/nms_cuda.cpp
new file mode 100644
index 0000000..0ea6f9b
--- /dev/null
+++ b/lib/ops/nms/src/nms_cuda.cpp
@@ -0,0 +1,17 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+#include <torch/extension.h>
+
+#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
+
+at::Tensor nms_cuda(const at::Tensor boxes, float nms_overlap_thresh);
+
+at::Tensor nms(const at::Tensor& dets, const float threshold) {
+  CHECK_CUDA(dets);
+  if (dets.numel() == 0)
+    return at::empty({0}, dets.options().dtype(at::kLong).device(at::kCPU));
+  return nms_cuda(dets, threshold);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("nms", &nms, "non-maximum suppression");
+}
\ No newline at end of file
diff --git a/lib/ops/nms/src/nms_kernel.cu b/lib/ops/nms/src/nms_kernel.cu
new file mode 100644
index 0000000..9254f2a
--- /dev/null
+++ b/lib/ops/nms/src/nms_kernel.cu
@@ -0,0 +1,131 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <THC/THC.h>
+#include <THC/THCDeviceUtils.cuh>
+
+#include <vector>
+#include <iostream>
+
+int const threadsPerBlock = sizeof(unsigned long long) * 8;
+
+__device__ inline float devIoU(float const * const a, float const * const b) {
+  float left = max(a[0], b[0]), right = min(a[2], b[2]);
+  float top = max(a[1], b[1]), bottom = min(a[3], b[3]);
+  float width = max(right - left + 1, 0.f), height = max(bottom - top + 1, 0.f);
+  float interS = width * height;
+  float Sa = (a[2] - a[0] + 1) * (a[3] - a[1] + 1);
+  float Sb = (b[2] - b[0] + 1) * (b[3] - b[1] + 1);
+  return interS / (Sa + Sb - interS);
+}
+
+__global__ void nms_kernel(const int n_boxes, const float nms_overlap_thresh,
+                           const float *dev_boxes, unsigned long long *dev_mask) {
+  const int row_start = blockIdx.y;
+  const int col_start = blockIdx.x;
+
+  // if (row_start > col_start) return;
+
+  const int row_size =
+        min(n_boxes - row_start * threadsPerBlock, threadsPerBlock);
+  const int col_size =
+        min(n_boxes - col_start * threadsPerBlock, threadsPerBlock);
+
+  __shared__ float block_boxes[threadsPerBlock * 5];
+  if (threadIdx.x < col_size) {
+    block_boxes[threadIdx.x * 5 + 0] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 0];
+    block_boxes[threadIdx.x * 5 + 1] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 1];
+    block_boxes[threadIdx.x * 5 + 2] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 2];
+    block_boxes[threadIdx.x * 5 + 3] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 3];
+    block_boxes[threadIdx.x * 5 + 4] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 4];
+  }
+  __syncthreads();
+
+  if (threadIdx.x < row_size) {
+    const int cur_box_idx = threadsPerBlock * row_start + threadIdx.x;
+    const float *cur_box = dev_boxes + cur_box_idx * 5;
+    int i = 0;
+    unsigned long long t = 0;
+    int start = 0;
+    if (row_start == col_start) {
+      start = threadIdx.x + 1;
+    }
+    for (i = start; i < col_size; i++) {
+      if (devIoU(cur_box, block_boxes + i * 5) > nms_overlap_thresh) {
+        t |= 1ULL << i;
+      }
+    }
+    const int col_blocks = THCCeilDiv(n_boxes, threadsPerBlock);
+    dev_mask[cur_box_idx * col_blocks + col_start] = t;
+  }
+}
+
+// boxes is a N x 5 tensor
+at::Tensor nms_cuda(const at::Tensor boxes, float nms_overlap_thresh) {
+  using scalar_t = float;
+  AT_ASSERTM(boxes.type().is_cuda(), "boxes must be a CUDA tensor");
+  auto scores = boxes.select(1, 4);
+  auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
+  auto boxes_sorted = boxes.index_select(0, order_t);
+
+  int boxes_num = boxes.size(0);
+
+  const int col_blocks = THCCeilDiv(boxes_num, threadsPerBlock);
+
+  scalar_t* boxes_dev = boxes_sorted.data<scalar_t>();
+
+  THCState *state = at::globalContext().lazyInitCUDA(); // TODO replace with getTHCState
+
+  unsigned long long* mask_dev = NULL;
+  //THCudaCheck(THCudaMalloc(state, (void**) &mask_dev,
+  //                      boxes_num * col_blocks * sizeof(unsigned long long)));
+
+  mask_dev = (unsigned long long*) THCudaMalloc(state, boxes_num * col_blocks * sizeof(unsigned long long));
+
+  dim3 blocks(THCCeilDiv(boxes_num, threadsPerBlock),
+              THCCeilDiv(boxes_num, threadsPerBlock));
+  dim3 threads(threadsPerBlock);
+  nms_kernel<<<blocks, threads>>>(boxes_num,
+                                  nms_overlap_thresh,
+                                  boxes_dev,
+                                  mask_dev);
+
+  std::vector<unsigned long long> mask_host(boxes_num * col_blocks);
+  THCudaCheck(cudaMemcpy(&mask_host[0],
+                        mask_dev,
+                        sizeof(unsigned long long) * boxes_num * col_blocks,
+                        cudaMemcpyDeviceToHost));
+
+  std::vector<unsigned long long> remv(col_blocks);
+  memset(&remv[0], 0, sizeof(unsigned long long) * col_blocks);
+
+  at::Tensor keep = at::empty({boxes_num}, boxes.options().dtype(at::kLong).device(at::kCPU));
+  int64_t* keep_out = keep.data<int64_t>();
+
+  int num_to_keep = 0;
+  for (int i = 0; i < boxes_num; i++) {
+    int nblock = i / threadsPerBlock;
+    int inblock = i % threadsPerBlock;
+
+    if (!(remv[nblock] & (1ULL << inblock))) {
+      keep_out[num_to_keep++] = i;
+      unsigned long long *p = &mask_host[0] + i * col_blocks;
+      for (int j = nblock; j < col_blocks; j++) {
+        remv[j] |= p[j];
+      }
+    }
+  }
+
+  THCudaFree(state, mask_dev);
+  // TODO improve this part
+  return std::get<0>(order_t.index({
+                       keep.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep).to(
+                         order_t.device(), keep.scalar_type())
+                     }).sort(0, false));
+}
\ No newline at end of file
diff --git a/lib/ops/nms/src/soft_nms_cpu.cpp b/lib/ops/nms/src/soft_nms_cpu.cpp
new file mode 100644
index 0000000..9615693
--- /dev/null
+++ b/lib/ops/nms/src/soft_nms_cpu.cpp
@@ -0,0 +1,10323 @@
+/* Generated by Cython 0.29.12 */
+
+/* BEGIN: Cython Metadata
+{
+    "distutils": {
+        "depends": [
+            "/home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/numpy/core/include/numpy/arrayobject.h",
+            "/home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/numpy/core/include/numpy/ufuncobject.h"
+        ],
+        "extra_compile_args": {
+            "cxx": [
+                "-Wno-unused-function",
+                "-Wno-write-strings"
+            ]
+        },
+        "include_dirs": [
+            "/home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/numpy/core/include"
+        ],
+        "language": "c++",
+        "name": "mmdet.ops.nms.soft_nms_cpu",
+        "sources": [
+            "mmdet/ops/nms/src/soft_nms_cpu.pyx"
+        ]
+    },
+    "module_name": "mmdet.ops.nms.soft_nms_cpu"
+}
+END: Cython Metadata */
+
+#define PY_SSIZE_T_CLEAN
+#include "Python.h"
+#ifndef Py_PYTHON_H
+    #error Python headers needed to compile C extensions, please install development version of Python.
+#elif PY_VERSION_HEX < 0x02060000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000)
+    #error Cython requires Python 2.6+ or Python 3.3+.
+#else
+#define CYTHON_ABI "0_29_12"
+#define CYTHON_HEX_VERSION 0x001D0CF0
+#define CYTHON_FUTURE_DIVISION 1
+#include <stddef.h>
+#ifndef offsetof
+  #define offsetof(type, member) ( (size_t) & ((type*)0) -> member )
+#endif
+#if !defined(WIN32) && !defined(MS_WINDOWS)
+  #ifndef __stdcall
+    #define __stdcall
+  #endif
+  #ifndef __cdecl
+    #define __cdecl
+  #endif
+  #ifndef __fastcall
+    #define __fastcall
+  #endif
+#endif
+#ifndef DL_IMPORT
+  #define DL_IMPORT(t) t
+#endif
+#ifndef DL_EXPORT
+  #define DL_EXPORT(t) t
+#endif
+#define __PYX_COMMA ,
+#ifndef HAVE_LONG_LONG
+  #if PY_VERSION_HEX >= 0x02070000
+    #define HAVE_LONG_LONG
+  #endif
+#endif
+#ifndef PY_LONG_LONG
+  #define PY_LONG_LONG LONG_LONG
+#endif
+#ifndef Py_HUGE_VAL
+  #define Py_HUGE_VAL HUGE_VAL
+#endif
+#ifdef PYPY_VERSION
+  #define CYTHON_COMPILING_IN_PYPY 1
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #undef CYTHON_USE_TYPE_SLOTS
+  #define CYTHON_USE_TYPE_SLOTS 0
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #if PY_VERSION_HEX < 0x03050000
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_INTERNALS
+  #define CYTHON_USE_UNICODE_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #undef CYTHON_AVOID_BORROWED_REFS
+  #define CYTHON_AVOID_BORROWED_REFS 1
+  #undef CYTHON_ASSUME_SAFE_MACROS
+  #define CYTHON_ASSUME_SAFE_MACROS 0
+  #undef CYTHON_UNPACK_METHODS
+  #define CYTHON_UNPACK_METHODS 0
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+#elif defined(PYSTON_VERSION)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 1
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #undef CYTHON_USE_ASYNC_SLOTS
+  #define CYTHON_USE_ASYNC_SLOTS 0
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+#else
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 1
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYTYPE_LOOKUP
+    #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #elif !defined(CYTHON_USE_PYTYPE_LOOKUP)
+    #define CYTHON_USE_PYTYPE_LOOKUP 1
+  #endif
+  #if PY_MAJOR_VERSION < 3
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYLONG_INTERNALS
+    #define CYTHON_USE_PYLONG_INTERNALS 0
+  #elif !defined(CYTHON_USE_PYLONG_INTERNALS)
+    #define CYTHON_USE_PYLONG_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_PYLIST_INTERNALS
+    #define CYTHON_USE_PYLIST_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #if PY_VERSION_HEX < 0x030300F0
+    #undef CYTHON_USE_UNICODE_WRITER
+    #define CYTHON_USE_UNICODE_WRITER 0
+  #elif !defined(CYTHON_USE_UNICODE_WRITER)
+    #define CYTHON_USE_UNICODE_WRITER 1
+  #endif
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #ifndef CYTHON_FAST_THREAD_STATE
+    #define CYTHON_FAST_THREAD_STATE 1
+  #endif
+  #ifndef CYTHON_FAST_PYCALL
+    #define CYTHON_FAST_PYCALL 1
+  #endif
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT (PY_VERSION_HEX >= 0x03050000)
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1)
+  #endif
+  #ifndef CYTHON_USE_DICT_VERSIONS
+    #define CYTHON_USE_DICT_VERSIONS (PY_VERSION_HEX >= 0x030600B1)
+  #endif
+  #ifndef CYTHON_USE_EXC_INFO_STACK
+    #define CYTHON_USE_EXC_INFO_STACK (PY_VERSION_HEX >= 0x030700A3)
+  #endif
+#endif
+#if !defined(CYTHON_FAST_PYCCALL)
+#define CYTHON_FAST_PYCCALL  (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1)
+#endif
+#if CYTHON_USE_PYLONG_INTERNALS
+  #include "longintrepr.h"
+  #undef SHIFT
+  #undef BASE
+  #undef MASK
+  #ifdef SIZEOF_VOID_P
+    enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) };
+  #endif
+#endif
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+#ifndef __has_cpp_attribute
+  #define __has_cpp_attribute(x) 0
+#endif
+#ifndef CYTHON_RESTRICT
+  #if defined(__GNUC__)
+    #define CYTHON_RESTRICT __restrict__
+  #elif defined(_MSC_VER) && _MSC_VER >= 1400
+    #define CYTHON_RESTRICT __restrict
+  #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define CYTHON_RESTRICT restrict
+  #else
+    #define CYTHON_RESTRICT
+  #endif
+#endif
+#ifndef CYTHON_UNUSED
+# if defined(__GNUC__)
+#   if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4))
+#     define CYTHON_UNUSED __attribute__ ((__unused__))
+#   else
+#     define CYTHON_UNUSED
+#   endif
+# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER))
+#   define CYTHON_UNUSED __attribute__ ((__unused__))
+# else
+#   define CYTHON_UNUSED
+# endif
+#endif
+#ifndef CYTHON_MAYBE_UNUSED_VAR
+#  if defined(__cplusplus)
+     template<class T> void CYTHON_MAYBE_UNUSED_VAR( const T& ) { }
+#  else
+#    define CYTHON_MAYBE_UNUSED_VAR(x) (void)(x)
+#  endif
+#endif
+#ifndef CYTHON_NCP_UNUSED
+# if CYTHON_COMPILING_IN_CPYTHON
+#  define CYTHON_NCP_UNUSED
+# else
+#  define CYTHON_NCP_UNUSED CYTHON_UNUSED
+# endif
+#endif
+#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None)
+#ifdef _MSC_VER
+    #ifndef _MSC_STDINT_H_
+        #if _MSC_VER < 1300
+           typedef unsigned char     uint8_t;
+           typedef unsigned int      uint32_t;
+        #else
+           typedef unsigned __int8   uint8_t;
+           typedef unsigned __int32  uint32_t;
+        #endif
+    #endif
+#else
+   #include <stdint.h>
+#endif
+#ifndef CYTHON_FALLTHROUGH
+  #if defined(__cplusplus) && __cplusplus >= 201103L
+    #if __has_cpp_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH [[fallthrough]]
+    #elif __has_cpp_attribute(clang::fallthrough)
+      #define CYTHON_FALLTHROUGH [[clang::fallthrough]]
+    #elif __has_cpp_attribute(gnu::fallthrough)
+      #define CYTHON_FALLTHROUGH [[gnu::fallthrough]]
+    #endif
+  #endif
+  #ifndef CYTHON_FALLTHROUGH
+    #if __has_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH __attribute__((fallthrough))
+    #else
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+  #if defined(__clang__ ) && defined(__apple_build_version__)
+    #if __apple_build_version__ < 7000000
+      #undef  CYTHON_FALLTHROUGH
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+#endif
+
+#ifndef __cplusplus
+  #error "Cython files generated with the C++ option must be compiled with a C++ compiler."
+#endif
+#ifndef CYTHON_INLINE
+  #if defined(__clang__)
+    #define CYTHON_INLINE __inline__ __attribute__ ((__unused__))
+  #else
+    #define CYTHON_INLINE inline
+  #endif
+#endif
+template<typename T>
+void __Pyx_call_destructor(T& x) {
+    x.~T();
+}
+template<typename T>
+class __Pyx_FakeReference {
+  public:
+    __Pyx_FakeReference() : ptr(NULL) { }
+    __Pyx_FakeReference(const T& ref) : ptr(const_cast<T*>(&ref)) { }
+    T *operator->() { return ptr; }
+    T *operator&() { return ptr; }
+    operator T&() { return *ptr; }
+    template<typename U> bool operator ==(U other) { return *ptr == other; }
+    template<typename U> bool operator !=(U other) { return *ptr != other; }
+  private:
+    T *ptr;
+};
+
+#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag)
+  #define Py_OptimizeFlag 0
+#endif
+#define __PYX_BUILD_PY_SSIZE_T "n"
+#define CYTHON_FORMAT_SSIZE_T "z"
+#if PY_MAJOR_VERSION < 3
+  #define __Pyx_BUILTIN_MODULE_NAME "__builtin__"
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+  #define __Pyx_DefaultClassType PyClass_Type
+#else
+  #define __Pyx_BUILTIN_MODULE_NAME "builtins"
+#if PY_VERSION_HEX >= 0x030800A4 && PY_VERSION_HEX < 0x030800B2
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, 0, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+#else
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+#endif
+  #define __Pyx_DefaultClassType PyType_Type
+#endif
+#ifndef Py_TPFLAGS_CHECKTYPES
+  #define Py_TPFLAGS_CHECKTYPES 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_INDEX
+  #define Py_TPFLAGS_HAVE_INDEX 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_NEWBUFFER
+  #define Py_TPFLAGS_HAVE_NEWBUFFER 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_FINALIZE
+  #define Py_TPFLAGS_HAVE_FINALIZE 0
+#endif
+#ifndef METH_STACKLESS
+  #define METH_STACKLESS 0
+#endif
+#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL)
+  #ifndef METH_FASTCALL
+     #define METH_FASTCALL 0x80
+  #endif
+  typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs);
+  typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args,
+                                                          Py_ssize_t nargs, PyObject *kwnames);
+#else
+  #define __Pyx_PyCFunctionFast _PyCFunctionFast
+  #define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords
+#endif
+#if CYTHON_FAST_PYCCALL
+#define __Pyx_PyFastCFunction_Check(func)\
+    ((PyCFunction_Check(func) && (METH_FASTCALL == (PyCFunction_GET_FLAGS(func) & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS | METH_STACKLESS)))))
+#else
+#define __Pyx_PyFastCFunction_Check(func) 0
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc)
+  #define PyObject_Malloc(s)   PyMem_Malloc(s)
+  #define PyObject_Free(p)     PyMem_Free(p)
+  #define PyObject_Realloc(p)  PyMem_Realloc(p)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030400A1
+  #define PyMem_RawMalloc(n)           PyMem_Malloc(n)
+  #define PyMem_RawRealloc(p, n)       PyMem_Realloc(p, n)
+  #define PyMem_RawFree(p)             PyMem_Free(p)
+#endif
+#if CYTHON_COMPILING_IN_PYSTON
+  #define __Pyx_PyCode_HasFreeVars(co)  PyCode_HasFreeVars(co)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno) PyFrame_SetLineNumber(frame, lineno)
+#else
+  #define __Pyx_PyCode_HasFreeVars(co)  (PyCode_GetNumFree(co) > 0)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno)  (frame)->f_lineno = (lineno)
+#endif
+#if !CYTHON_FAST_THREAD_STATE || PY_VERSION_HEX < 0x02070000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#elif PY_VERSION_HEX >= 0x03060000
+  #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet()
+#elif PY_VERSION_HEX >= 0x03000000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#else
+  #define __Pyx_PyThreadState_Current _PyThreadState_Current
+#endif
+#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT)
+#include "pythread.h"
+#define Py_tss_NEEDS_INIT 0
+typedef int Py_tss_t;
+static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) {
+  *key = PyThread_create_key();
+  return 0;
+}
+static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) {
+  Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t));
+  *key = Py_tss_NEEDS_INIT;
+  return key;
+}
+static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) {
+  PyObject_Free(key);
+}
+static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) {
+  return *key != Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) {
+  PyThread_delete_key(*key);
+  *key = Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) {
+  return PyThread_set_key_value(*key, value);
+}
+static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) {
+  return PyThread_get_key_value(*key);
+}
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON || defined(_PyDict_NewPresized)
+#define __Pyx_PyDict_NewPresized(n)  ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n))
+#else
+#define __Pyx_PyDict_NewPresized(n)  PyDict_New()
+#endif
+#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_TrueDivide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceTrueDivide(x,y)
+#else
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_Divide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceDivide(x,y)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && CYTHON_USE_UNICODE_INTERNALS
+#define __Pyx_PyDict_GetItemStr(dict, name)  _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash)
+#else
+#define __Pyx_PyDict_GetItemStr(dict, name)  PyDict_GetItem(dict, name)
+#endif
+#if PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND)
+  #define CYTHON_PEP393_ENABLED 1
+  #define __Pyx_PyUnicode_READY(op)       (likely(PyUnicode_IS_READY(op)) ?\
+                                              0 : _PyUnicode_Ready((PyObject *)(op)))
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_LENGTH(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i)
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   PyUnicode_MAX_CHAR_VALUE(u)
+  #define __Pyx_PyUnicode_KIND(u)         PyUnicode_KIND(u)
+  #define __Pyx_PyUnicode_DATA(u)         PyUnicode_DATA(u)
+  #define __Pyx_PyUnicode_READ(k, d, i)   PyUnicode_READ(k, d, i)
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  PyUnicode_WRITE(k, d, i, ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u)))
+#else
+  #define CYTHON_PEP393_ENABLED 0
+  #define PyUnicode_1BYTE_KIND  1
+  #define PyUnicode_2BYTE_KIND  2
+  #define PyUnicode_4BYTE_KIND  4
+  #define __Pyx_PyUnicode_READY(op)       (0)
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_SIZE(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i]))
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   ((sizeof(Py_UNICODE) == 2) ? 65535 : 1114111)
+  #define __Pyx_PyUnicode_KIND(u)         (sizeof(Py_UNICODE))
+  #define __Pyx_PyUnicode_DATA(u)         ((void*)PyUnicode_AS_UNICODE(u))
+  #define __Pyx_PyUnicode_READ(k, d, i)   ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i]))
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  (((void)(k)), ((Py_UNICODE*)d)[i] = ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_SIZE(u))
+#endif
+#if CYTHON_COMPILING_IN_PYPY
+  #define __Pyx_PyUnicode_Concat(a, b)      PyNumber_Add(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  PyNumber_Add(a, b)
+#else
+  #define __Pyx_PyUnicode_Concat(a, b)      PyUnicode_Concat(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\
+      PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b))
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyUnicode_Contains)
+  #define PyUnicode_Contains(u, s)  PySequence_Contains(u, s)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyByteArray_Check)
+  #define PyByteArray_Check(obj)  PyObject_TypeCheck(obj, &PyByteArray_Type)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Format)
+  #define PyObject_Format(obj, fmt)  PyObject_CallMethod(obj, "__format__", "O", fmt)
+#endif
+#define __Pyx_PyString_FormatSafe(a, b)   ((unlikely((a) == Py_None || (PyString_Check(b) && !PyString_CheckExact(b)))) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b))
+#define __Pyx_PyUnicode_FormatSafe(a, b)  ((unlikely((a) == Py_None || (PyUnicode_Check(b) && !PyUnicode_CheckExact(b)))) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b))
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyString_Format(a, b)  PyUnicode_Format(a, b)
+#else
+  #define __Pyx_PyString_Format(a, b)  PyString_Format(a, b)
+#endif
+#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII)
+  #define PyObject_ASCII(o)            PyObject_Repr(o)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBaseString_Type            PyUnicode_Type
+  #define PyStringObject               PyUnicodeObject
+  #define PyString_Type                PyUnicode_Type
+  #define PyString_Check               PyUnicode_Check
+  #define PyString_CheckExact          PyUnicode_CheckExact
+  #define PyObject_Unicode             PyObject_Str
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj)
+  #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj)
+#else
+  #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj))
+  #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj))
+#endif
+#ifndef PySet_CheckExact
+  #define PySet_CheckExact(obj)        (Py_TYPE(obj) == &PySet_Type)
+#endif
+#if CYTHON_ASSUME_SAFE_MACROS
+  #define __Pyx_PySequence_SIZE(seq)  Py_SIZE(seq)
+#else
+  #define __Pyx_PySequence_SIZE(seq)  PySequence_Size(seq)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyIntObject                  PyLongObject
+  #define PyInt_Type                   PyLong_Type
+  #define PyInt_Check(op)              PyLong_Check(op)
+  #define PyInt_CheckExact(op)         PyLong_CheckExact(op)
+  #define PyInt_FromString             PyLong_FromString
+  #define PyInt_FromUnicode            PyLong_FromUnicode
+  #define PyInt_FromLong               PyLong_FromLong
+  #define PyInt_FromSize_t             PyLong_FromSize_t
+  #define PyInt_FromSsize_t            PyLong_FromSsize_t
+  #define PyInt_AsLong                 PyLong_AsLong
+  #define PyInt_AS_LONG                PyLong_AS_LONG
+  #define PyInt_AsSsize_t              PyLong_AsSsize_t
+  #define PyInt_AsUnsignedLongMask     PyLong_AsUnsignedLongMask
+  #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask
+  #define PyNumber_Int                 PyNumber_Long
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBoolObject                 PyLongObject
+#endif
+#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY
+  #ifndef PyUnicode_InternFromString
+    #define PyUnicode_InternFromString(s) PyUnicode_FromString(s)
+  #endif
+#endif
+#if PY_VERSION_HEX < 0x030200A4
+  typedef long Py_hash_t;
+  #define __Pyx_PyInt_FromHash_t PyInt_FromLong
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsLong
+#else
+  #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsSsize_t
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyMethod_New(func, self, klass) ((self) ? PyMethod_New(func, self) : (Py_INCREF(func), func))
+#else
+  #define __Pyx_PyMethod_New(func, self, klass) PyMethod_New(func, self, klass)
+#endif
+#if CYTHON_USE_ASYNC_SLOTS
+  #if PY_VERSION_HEX >= 0x030500B1
+    #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods
+    #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async)
+  #else
+    #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved))
+  #endif
+#else
+  #define __Pyx_PyType_AsAsync(obj) NULL
+#endif
+#ifndef __Pyx_PyAsyncMethodsStruct
+    typedef struct {
+        unaryfunc am_await;
+        unaryfunc am_aiter;
+        unaryfunc am_anext;
+    } __Pyx_PyAsyncMethodsStruct;
+#endif
+
+#if defined(WIN32) || defined(MS_WINDOWS)
+  #define _USE_MATH_DEFINES
+#endif
+#include <math.h>
+#ifdef NAN
+#define __PYX_NAN() ((float) NAN)
+#else
+static CYTHON_INLINE float __PYX_NAN() {
+  float value;
+  memset(&value, 0xFF, sizeof(value));
+  return value;
+}
+#endif
+#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL)
+#define __Pyx_truncl trunc
+#else
+#define __Pyx_truncl truncl
+#endif
+
+
+#define __PYX_ERR(f_index, lineno, Ln_error) \
+{ \
+  __pyx_filename = __pyx_f[f_index]; __pyx_lineno = lineno; __pyx_clineno = __LINE__; goto Ln_error; \
+}
+
+#ifndef __PYX_EXTERN_C
+  #ifdef __cplusplus
+    #define __PYX_EXTERN_C extern "C"
+  #else
+    #define __PYX_EXTERN_C extern
+  #endif
+#endif
+
+#define __PYX_HAVE__mmdet__ops__nms__soft_nms_cpu
+#define __PYX_HAVE_API__mmdet__ops__nms__soft_nms_cpu
+/* Early includes */
+#include <string.h>
+#include <stdio.h>
+#include "numpy/arrayobject.h"
+#include "numpy/ufuncobject.h"
+#ifdef _OPENMP
+#include <omp.h>
+#endif /* _OPENMP */
+
+#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS)
+#define CYTHON_WITHOUT_ASSERTIONS
+#endif
+
+typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding;
+                const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry;
+
+#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_UTF8 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT (PY_MAJOR_VERSION >= 3 && __PYX_DEFAULT_STRING_ENCODING_IS_UTF8)
+#define __PYX_DEFAULT_STRING_ENCODING ""
+#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString
+#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#define __Pyx_uchar_cast(c) ((unsigned char)c)
+#define __Pyx_long_cast(x) ((long)x)
+#define __Pyx_fits_Py_ssize_t(v, type, is_signed)  (\
+    (sizeof(type) < sizeof(Py_ssize_t))  ||\
+    (sizeof(type) > sizeof(Py_ssize_t) &&\
+          likely(v < (type)PY_SSIZE_T_MAX ||\
+                 v == (type)PY_SSIZE_T_MAX)  &&\
+          (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\
+                                v == (type)PY_SSIZE_T_MIN)))  ||\
+    (sizeof(type) == sizeof(Py_ssize_t) &&\
+          (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\
+                               v == (type)PY_SSIZE_T_MAX)))  )
+static CYTHON_INLINE int __Pyx_is_valid_index(Py_ssize_t i, Py_ssize_t limit) {
+    return (size_t) i < (size_t) limit;
+}
+#if defined (__cplusplus) && __cplusplus >= 201103L
+    #include <cstdlib>
+    #define __Pyx_sst_abs(value) std::abs(value)
+#elif SIZEOF_INT >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) abs(value)
+#elif SIZEOF_LONG >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) labs(value)
+#elif defined (_MSC_VER)
+    #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value))
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define __Pyx_sst_abs(value) llabs(value)
+#elif defined (__GNUC__)
+    #define __Pyx_sst_abs(value) __builtin_llabs(value)
+#else
+    #define __Pyx_sst_abs(value) ((value<0) ? -value : value)
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*);
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length);
+#define __Pyx_PyByteArray_FromString(s) PyByteArray_FromStringAndSize((const char*)s, strlen((const char*)s))
+#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l)
+#define __Pyx_PyBytes_FromString        PyBytes_FromString
+#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*);
+#if PY_MAJOR_VERSION < 3
+    #define __Pyx_PyStr_FromString        __Pyx_PyBytes_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#else
+    #define __Pyx_PyStr_FromString        __Pyx_PyUnicode_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize
+#endif
+#define __Pyx_PyBytes_AsWritableString(s)     ((char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableSString(s)    ((signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableUString(s)    ((unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsString(s)     ((const char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsSString(s)    ((const signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsUString(s)    ((const unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyObject_AsWritableString(s)    ((char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableSString(s)    ((signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableUString(s)    ((unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsSString(s)    ((const signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsUString(s)    ((const unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_FromCString(s)  __Pyx_PyObject_FromString((const char*)s)
+#define __Pyx_PyBytes_FromCString(s)   __Pyx_PyBytes_FromString((const char*)s)
+#define __Pyx_PyByteArray_FromCString(s)   __Pyx_PyByteArray_FromString((const char*)s)
+#define __Pyx_PyStr_FromCString(s)     __Pyx_PyStr_FromString((const char*)s)
+#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s)
+static CYTHON_INLINE size_t __Pyx_Py_UNICODE_strlen(const Py_UNICODE *u) {
+    const Py_UNICODE *u_end = u;
+    while (*u_end++) ;
+    return (size_t)(u_end - u - 1);
+}
+#define __Pyx_PyUnicode_FromUnicode(u)       PyUnicode_FromUnicode(u, __Pyx_Py_UNICODE_strlen(u))
+#define __Pyx_PyUnicode_FromUnicodeAndLength PyUnicode_FromUnicode
+#define __Pyx_PyUnicode_AsUnicode            PyUnicode_AsUnicode
+#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj)
+#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None)
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject*);
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x);
+#define __Pyx_PySequence_Tuple(obj)\
+    (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj))
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*);
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t);
+#if CYTHON_ASSUME_SAFE_MACROS
+#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x))
+#else
+#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x)
+#endif
+#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x))
+#if PY_MAJOR_VERSION >= 3
+#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x))
+#else
+#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x))
+#endif
+#define __Pyx_PyNumber_Float(x) (PyFloat_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Float(x))
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+static int __Pyx_sys_getdefaultencoding_not_ascii;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    PyObject* ascii_chars_u = NULL;
+    PyObject* ascii_chars_b = NULL;
+    const char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    if (strcmp(default_encoding_c, "ascii") == 0) {
+        __Pyx_sys_getdefaultencoding_not_ascii = 0;
+    } else {
+        char ascii_chars[128];
+        int c;
+        for (c = 0; c < 128; c++) {
+            ascii_chars[c] = c;
+        }
+        __Pyx_sys_getdefaultencoding_not_ascii = 1;
+        ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL);
+        if (!ascii_chars_u) goto bad;
+        ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL);
+        if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.",
+                default_encoding_c);
+            goto bad;
+        }
+        Py_DECREF(ascii_chars_u);
+        Py_DECREF(ascii_chars_b);
+    }
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    Py_XDECREF(ascii_chars_u);
+    Py_XDECREF(ascii_chars_b);
+    return -1;
+}
+#endif
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL)
+#else
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL)
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+static char* __PYX_DEFAULT_STRING_ENCODING;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c) + 1);
+    if (!__PYX_DEFAULT_STRING_ENCODING) goto bad;
+    strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c);
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    return -1;
+}
+#endif
+#endif
+
+
+/* Test for GCC > 2.95 */
+#if defined(__GNUC__)     && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95)))
+  #define likely(x)   __builtin_expect(!!(x), 1)
+  #define unlikely(x) __builtin_expect(!!(x), 0)
+#else /* !__GNUC__ or GCC < 2.95 */
+  #define likely(x)   (x)
+  #define unlikely(x) (x)
+#endif /* __GNUC__ */
+static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; }
+
+static PyObject *__pyx_m = NULL;
+static PyObject *__pyx_d;
+static PyObject *__pyx_b;
+static PyObject *__pyx_cython_runtime = NULL;
+static PyObject *__pyx_empty_tuple;
+static PyObject *__pyx_empty_bytes;
+static PyObject *__pyx_empty_unicode;
+static int __pyx_lineno;
+static int __pyx_clineno = 0;
+static const char * __pyx_cfilenm= __FILE__;
+static const char *__pyx_filename;
+
+/* Header.proto */
+#if !defined(CYTHON_CCOMPLEX)
+  #if defined(__cplusplus)
+    #define CYTHON_CCOMPLEX 1
+  #elif defined(_Complex_I)
+    #define CYTHON_CCOMPLEX 1
+  #else
+    #define CYTHON_CCOMPLEX 0
+  #endif
+#endif
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #include <complex>
+  #else
+    #include <complex.h>
+  #endif
+#endif
+#if CYTHON_CCOMPLEX && !defined(__cplusplus) && defined(__sun__) && defined(__GNUC__)
+  #undef _Complex_I
+  #define _Complex_I 1.0fj
+#endif
+
+
+static const char *__pyx_f[] = {
+  "mmdet/ops/nms/src/soft_nms_cpu.pyx",
+  "__init__.pxd",
+  "type.pxd",
+};
+/* BufferFormatStructs.proto */
+#define IS_UNSIGNED(type) (((type) -1) > 0)
+struct __Pyx_StructField_;
+#define __PYX_BUF_FLAGS_PACKED_STRUCT (1 << 0)
+typedef struct {
+  const char* name;
+  struct __Pyx_StructField_* fields;
+  size_t size;
+  size_t arraysize[8];
+  int ndim;
+  char typegroup;
+  char is_unsigned;
+  int flags;
+} __Pyx_TypeInfo;
+typedef struct __Pyx_StructField_ {
+  __Pyx_TypeInfo* type;
+  const char* name;
+  size_t offset;
+} __Pyx_StructField;
+typedef struct {
+  __Pyx_StructField* field;
+  size_t parent_offset;
+} __Pyx_BufFmt_StackElem;
+typedef struct {
+  __Pyx_StructField root;
+  __Pyx_BufFmt_StackElem* head;
+  size_t fmt_offset;
+  size_t new_count, enc_count;
+  size_t struct_alignment;
+  int is_complex;
+  char enc_type;
+  char new_packmode;
+  char enc_packmode;
+  char is_valid_array;
+} __Pyx_BufFmt_Context;
+
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":776
+ * # in Cython to enable them only on the right systems.
+ * 
+ * ctypedef npy_int8       int8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ */
+typedef npy_int8 __pyx_t_5numpy_int8_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":777
+ * 
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t
+ */
+typedef npy_int16 __pyx_t_5numpy_int16_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":778
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int64      int64_t
+ * #ctypedef npy_int96      int96_t
+ */
+typedef npy_int32 __pyx_t_5numpy_int32_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":779
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_int96      int96_t
+ * #ctypedef npy_int128     int128_t
+ */
+typedef npy_int64 __pyx_t_5numpy_int64_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":783
+ * #ctypedef npy_int128     int128_t
+ * 
+ * ctypedef npy_uint8      uint8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ */
+typedef npy_uint8 __pyx_t_5numpy_uint8_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":784
+ * 
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t
+ */
+typedef npy_uint16 __pyx_t_5numpy_uint16_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":785
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint64     uint64_t
+ * #ctypedef npy_uint96     uint96_t
+ */
+typedef npy_uint32 __pyx_t_5numpy_uint32_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":786
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_uint96     uint96_t
+ * #ctypedef npy_uint128    uint128_t
+ */
+typedef npy_uint64 __pyx_t_5numpy_uint64_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":790
+ * #ctypedef npy_uint128    uint128_t
+ * 
+ * ctypedef npy_float32    float32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_float64    float64_t
+ * #ctypedef npy_float80    float80_t
+ */
+typedef npy_float32 __pyx_t_5numpy_float32_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":791
+ * 
+ * ctypedef npy_float32    float32_t
+ * ctypedef npy_float64    float64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_float80    float80_t
+ * #ctypedef npy_float128   float128_t
+ */
+typedef npy_float64 __pyx_t_5numpy_float64_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":800
+ * # The int types are mapped a bit surprising --
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t
+ */
+typedef npy_long __pyx_t_5numpy_int_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":801
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   longlong_t
+ * 
+ */
+typedef npy_longlong __pyx_t_5numpy_long_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":802
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_ulong      uint_t
+ */
+typedef npy_longlong __pyx_t_5numpy_longlong_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":804
+ * ctypedef npy_longlong   longlong_t
+ * 
+ * ctypedef npy_ulong      uint_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t
+ */
+typedef npy_ulong __pyx_t_5numpy_uint_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":805
+ * 
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulong_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":806
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_intp       intp_t
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulonglong_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":808
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ * ctypedef npy_intp       intp_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uintp      uintp_t
+ * 
+ */
+typedef npy_intp __pyx_t_5numpy_intp_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":809
+ * 
+ * ctypedef npy_intp       intp_t
+ * ctypedef npy_uintp      uintp_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_double     float_t
+ */
+typedef npy_uintp __pyx_t_5numpy_uintp_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":811
+ * ctypedef npy_uintp      uintp_t
+ * 
+ * ctypedef npy_double     float_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t
+ */
+typedef npy_double __pyx_t_5numpy_float_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":812
+ * 
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ */
+typedef npy_double __pyx_t_5numpy_double_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":813
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ */
+typedef npy_longdouble __pyx_t_5numpy_longdouble_t;
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< float > __pyx_t_float_complex;
+  #else
+    typedef float _Complex __pyx_t_float_complex;
+  #endif
+#else
+    typedef struct { float real, imag; } __pyx_t_float_complex;
+#endif
+static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float, float);
+
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< double > __pyx_t_double_complex;
+  #else
+    typedef double _Complex __pyx_t_double_complex;
+  #endif
+#else
+    typedef struct { double real, imag; } __pyx_t_double_complex;
+#endif
+static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double, double);
+
+
+/*--- Type declarations ---*/
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":815
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ * ctypedef npy_cfloat      cfloat_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t
+ */
+typedef npy_cfloat __pyx_t_5numpy_cfloat_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":816
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ */
+typedef npy_cdouble __pyx_t_5numpy_cdouble_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":817
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cdouble     complex_t
+ */
+typedef npy_clongdouble __pyx_t_5numpy_clongdouble_t;
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":819
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ * ctypedef npy_cdouble     complex_t             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ */
+typedef npy_cdouble __pyx_t_5numpy_complex_t;
+
+/* --- Runtime support code (head) --- */
+/* Refnanny.proto */
+#ifndef CYTHON_REFNANNY
+  #define CYTHON_REFNANNY 0
+#endif
+#if CYTHON_REFNANNY
+  typedef struct {
+    void (*INCREF)(void*, PyObject*, int);
+    void (*DECREF)(void*, PyObject*, int);
+    void (*GOTREF)(void*, PyObject*, int);
+    void (*GIVEREF)(void*, PyObject*, int);
+    void* (*SetupContext)(const char*, int, const char*);
+    void (*FinishContext)(void**);
+  } __Pyx_RefNannyAPIStruct;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname);
+  #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL;
+#ifdef WITH_THREAD
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          if (acquire_gil) {\
+              PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+              PyGILState_Release(__pyx_gilstate_save);\
+          } else {\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+          }
+#else
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__)
+#endif
+  #define __Pyx_RefNannyFinishContext()\
+          __Pyx_RefNanny->FinishContext(&__pyx_refnanny)
+  #define __Pyx_INCREF(r)  __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_DECREF(r)  __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GOTREF(r)  __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_XINCREF(r)  do { if((r) != NULL) {__Pyx_INCREF(r); }} while(0)
+  #define __Pyx_XDECREF(r)  do { if((r) != NULL) {__Pyx_DECREF(r); }} while(0)
+  #define __Pyx_XGOTREF(r)  do { if((r) != NULL) {__Pyx_GOTREF(r); }} while(0)
+  #define __Pyx_XGIVEREF(r) do { if((r) != NULL) {__Pyx_GIVEREF(r);}} while(0)
+#else
+  #define __Pyx_RefNannyDeclarations
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)
+  #define __Pyx_RefNannyFinishContext()
+  #define __Pyx_INCREF(r) Py_INCREF(r)
+  #define __Pyx_DECREF(r) Py_DECREF(r)
+  #define __Pyx_GOTREF(r)
+  #define __Pyx_GIVEREF(r)
+  #define __Pyx_XINCREF(r) Py_XINCREF(r)
+  #define __Pyx_XDECREF(r) Py_XDECREF(r)
+  #define __Pyx_XGOTREF(r)
+  #define __Pyx_XGIVEREF(r)
+#endif
+#define __Pyx_XDECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_XDECREF(tmp);\
+    } while (0)
+#define __Pyx_DECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_DECREF(tmp);\
+    } while (0)
+#define __Pyx_CLEAR(r)    do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0)
+#define __Pyx_XCLEAR(r)   do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0)
+
+/* PyObjectGetAttrStr.proto */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name);
+#else
+#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n)
+#endif
+
+/* GetBuiltinName.proto */
+static PyObject *__Pyx_GetBuiltinName(PyObject *name);
+
+/* RaiseArgTupleInvalid.proto */
+static void __Pyx_RaiseArgtupleInvalid(const char* func_name, int exact,
+    Py_ssize_t num_min, Py_ssize_t num_max, Py_ssize_t num_found);
+
+/* RaiseDoubleKeywords.proto */
+static void __Pyx_RaiseDoubleKeywordsError(const char* func_name, PyObject* kw_name);
+
+/* ParseKeywords.proto */
+static int __Pyx_ParseOptionalKeywords(PyObject *kwds, PyObject **argnames[],\
+    PyObject *kwds2, PyObject *values[], Py_ssize_t num_pos_args,\
+    const char* function_name);
+
+/* ArgTypeTest.proto */
+#define __Pyx_ArgTypeTest(obj, type, none_allowed, name, exact)\
+    ((likely((Py_TYPE(obj) == type) | (none_allowed && (obj == Py_None)))) ? 1 :\
+        __Pyx__ArgTypeTest(obj, type, name, exact))
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact);
+
+/* IsLittleEndian.proto */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void);
+
+/* BufferFormatCheck.proto */
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts);
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type);
+
+/* BufferGetAndValidate.proto */
+#define __Pyx_GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)\
+    ((obj == Py_None || obj == NULL) ?\
+    (__Pyx_ZeroBuffer(buf), 0) :\
+    __Pyx__GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack))
+static int  __Pyx__GetBufferAndValidate(Py_buffer* buf, PyObject* obj,
+    __Pyx_TypeInfo* dtype, int flags, int nd, int cast, __Pyx_BufFmt_StackElem* stack);
+static void __Pyx_ZeroBuffer(Py_buffer* buf);
+static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info);
+static Py_ssize_t __Pyx_minusones[] = { -1, -1, -1, -1, -1, -1, -1, -1 };
+static Py_ssize_t __Pyx_zeros[] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+
+/* PyFunctionFastCall.proto */
+#if CYTHON_FAST_PYCALL
+#define __Pyx_PyFunction_FastCall(func, args, nargs)\
+    __Pyx_PyFunction_FastCallDict((func), (args), (nargs), NULL)
+#if 1 || PY_VERSION_HEX < 0x030600B1
+static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs);
+#else
+#define __Pyx_PyFunction_FastCallDict(func, args, nargs, kwargs) _PyFunction_FastCallDict(func, args, nargs, kwargs)
+#endif
+#define __Pyx_BUILD_ASSERT_EXPR(cond)\
+    (sizeof(char [1 - 2*!(cond)]) - 1)
+#ifndef Py_MEMBER_SIZE
+#define Py_MEMBER_SIZE(type, member) sizeof(((type *)0)->member)
+#endif
+  static size_t __pyx_pyframe_localsplus_offset = 0;
+  #include "frameobject.h"
+  #define __Pxy_PyFrame_Initialize_Offsets()\
+    ((void)__Pyx_BUILD_ASSERT_EXPR(sizeof(PyFrameObject) == offsetof(PyFrameObject, f_localsplus) + Py_MEMBER_SIZE(PyFrameObject, f_localsplus)),\
+     (void)(__pyx_pyframe_localsplus_offset = ((size_t)PyFrame_Type.tp_basicsize) - Py_MEMBER_SIZE(PyFrameObject, f_localsplus)))
+  #define __Pyx_PyFrame_GetLocalsplus(frame)\
+    (assert(__pyx_pyframe_localsplus_offset), (PyObject **)(((char *)(frame)) + __pyx_pyframe_localsplus_offset))
+#endif
+
+/* PyObjectCall.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw);
+#else
+#define __Pyx_PyObject_Call(func, arg, kw) PyObject_Call(func, arg, kw)
+#endif
+
+/* PyObjectCallMethO.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg);
+#endif
+
+/* PyObjectCallNoArg.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func);
+#else
+#define __Pyx_PyObject_CallNoArg(func) __Pyx_PyObject_Call(func, __pyx_empty_tuple, NULL)
+#endif
+
+/* PyCFunctionFastCall.proto */
+#if CYTHON_FAST_PYCCALL
+static CYTHON_INLINE PyObject *__Pyx_PyCFunction_FastCall(PyObject *func, PyObject **args, Py_ssize_t nargs);
+#else
+#define __Pyx_PyCFunction_FastCall(func, args, nargs)  (assert(0), NULL)
+#endif
+
+/* PyObjectCallOneArg.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg);
+
+/* GetItemInt.proto */
+#define __Pyx_GetItemInt(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\
+    (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\
+    __Pyx_GetItemInt_Fast(o, (Py_ssize_t)i, is_list, wraparound, boundscheck) :\
+    (is_list ? (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL) :\
+               __Pyx_GetItemInt_Generic(o, to_py_func(i))))
+#define __Pyx_GetItemInt_List(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\
+    (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\
+    __Pyx_GetItemInt_List_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\
+    (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL))
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i,
+                                                              int wraparound, int boundscheck);
+#define __Pyx_GetItemInt_Tuple(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\
+    (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\
+    __Pyx_GetItemInt_Tuple_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\
+    (PyErr_SetString(PyExc_IndexError, "tuple index out of range"), (PyObject*)NULL))
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i,
+                                                              int wraparound, int boundscheck);
+static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j);
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i,
+                                                     int is_list, int wraparound, int boundscheck);
+
+/* PyDictVersioning.proto */
+#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+#define __PYX_DICT_VERSION_INIT  ((PY_UINT64_T) -1)
+#define __PYX_GET_DICT_VERSION(dict)  (((PyDictObject*)(dict))->ma_version_tag)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)\
+    (version_var) = __PYX_GET_DICT_VERSION(dict);\
+    (cache_var) = (value);
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) {\
+    static PY_UINT64_T __pyx_dict_version = 0;\
+    static PyObject *__pyx_dict_cached_value = NULL;\
+    if (likely(__PYX_GET_DICT_VERSION(DICT) == __pyx_dict_version)) {\
+        (VAR) = __pyx_dict_cached_value;\
+    } else {\
+        (VAR) = __pyx_dict_cached_value = (LOOKUP);\
+        __pyx_dict_version = __PYX_GET_DICT_VERSION(DICT);\
+    }\
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj);
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj);
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version);
+#else
+#define __PYX_GET_DICT_VERSION(dict)  (0)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP)  (VAR) = (LOOKUP);
+#endif
+
+/* GetModuleGlobalName.proto */
+#if CYTHON_USE_DICT_VERSIONS
+#define __Pyx_GetModuleGlobalName(var, name)  {\
+    static PY_UINT64_T __pyx_dict_version = 0;\
+    static PyObject *__pyx_dict_cached_value = NULL;\
+    (var) = (likely(__pyx_dict_version == __PYX_GET_DICT_VERSION(__pyx_d))) ?\
+        (likely(__pyx_dict_cached_value) ? __Pyx_NewRef(__pyx_dict_cached_value) : __Pyx_GetBuiltinName(name)) :\
+        __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\
+}
+#define __Pyx_GetModuleGlobalNameUncached(var, name)  {\
+    PY_UINT64_T __pyx_dict_version;\
+    PyObject *__pyx_dict_cached_value;\
+    (var) = __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\
+}
+static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value);
+#else
+#define __Pyx_GetModuleGlobalName(var, name)  (var) = __Pyx__GetModuleGlobalName(name)
+#define __Pyx_GetModuleGlobalNameUncached(var, name)  (var) = __Pyx__GetModuleGlobalName(name)
+static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name);
+#endif
+
+/* PyObjectCall2Args.proto */
+static CYTHON_UNUSED PyObject* __Pyx_PyObject_Call2Args(PyObject* function, PyObject* arg1, PyObject* arg2);
+
+/* ObjectGetItem.proto */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject* key);
+#else
+#define __Pyx_PyObject_GetItem(obj, key)  PyObject_GetItem(obj, key)
+#endif
+
+/* PyIntBinop.proto */
+#if !CYTHON_COMPILING_IN_PYPY
+static PyObject* __Pyx_PyInt_AddObjC(PyObject *op1, PyObject *op2, long intval, int inplace, int zerodivision_check);
+#else
+#define __Pyx_PyInt_AddObjC(op1, op2, intval, inplace, zerodivision_check)\
+    (inplace ? PyNumber_InPlaceAdd(op1, op2) : PyNumber_Add(op1, op2))
+#endif
+
+/* SetItemInt.proto */
+#define __Pyx_SetItemInt(o, i, v, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\
+    (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\
+    __Pyx_SetItemInt_Fast(o, (Py_ssize_t)i, v, is_list, wraparound, boundscheck) :\
+    (is_list ? (PyErr_SetString(PyExc_IndexError, "list assignment index out of range"), -1) :\
+               __Pyx_SetItemInt_Generic(o, to_py_func(i), v)))
+static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v);
+static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v,
+                                               int is_list, int wraparound, int boundscheck);
+
+/* SliceObject.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetSlice(
+        PyObject* obj, Py_ssize_t cstart, Py_ssize_t cstop,
+        PyObject** py_start, PyObject** py_stop, PyObject** py_slice,
+        int has_cstart, int has_cstop, int wraparound);
+
+/* PyThreadStateGet.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyThreadState_declare  PyThreadState *__pyx_tstate;
+#define __Pyx_PyThreadState_assign  __pyx_tstate = __Pyx_PyThreadState_Current;
+#define __Pyx_PyErr_Occurred()  __pyx_tstate->curexc_type
+#else
+#define __Pyx_PyThreadState_declare
+#define __Pyx_PyThreadState_assign
+#define __Pyx_PyErr_Occurred()  PyErr_Occurred()
+#endif
+
+/* PyErrFetchRestore.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)    __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)    __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL))
+#else
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#endif
+#else
+#define __Pyx_PyErr_Clear() PyErr_Clear()
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestoreInState(tstate, type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchInState(tstate, type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)  PyErr_Fetch(type, value, tb)
+#endif
+
+/* RaiseException.proto */
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause);
+
+/* DictGetItem.proto */
+#if PY_MAJOR_VERSION >= 3 && !CYTHON_COMPILING_IN_PYPY
+static PyObject *__Pyx_PyDict_GetItem(PyObject *d, PyObject* key);
+#define __Pyx_PyObject_Dict_GetItem(obj, name)\
+    (likely(PyDict_CheckExact(obj)) ?\
+     __Pyx_PyDict_GetItem(obj, name) : PyObject_GetItem(obj, name))
+#else
+#define __Pyx_PyDict_GetItem(d, key) PyObject_GetItem(d, key)
+#define __Pyx_PyObject_Dict_GetItem(obj, name)  PyObject_GetItem(obj, name)
+#endif
+
+/* RaiseTooManyValuesToUnpack.proto */
+static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected);
+
+/* RaiseNeedMoreValuesToUnpack.proto */
+static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index);
+
+/* RaiseNoneIterError.proto */
+static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void);
+
+/* ExtTypeTest.proto */
+static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type);
+
+/* GetTopmostException.proto */
+#if CYTHON_USE_EXC_INFO_STACK
+static _PyErr_StackItem * __Pyx_PyErr_GetTopmostException(PyThreadState *tstate);
+#endif
+
+/* SaveResetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_ExceptionSave(type, value, tb)  __Pyx__ExceptionSave(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#define __Pyx_ExceptionReset(type, value, tb)  __Pyx__ExceptionReset(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+#else
+#define __Pyx_ExceptionSave(type, value, tb)   PyErr_GetExcInfo(type, value, tb)
+#define __Pyx_ExceptionReset(type, value, tb)  PyErr_SetExcInfo(type, value, tb)
+#endif
+
+/* PyErrExceptionMatches.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_ExceptionMatches(err) __Pyx_PyErr_ExceptionMatchesInState(__pyx_tstate, err)
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err);
+#else
+#define __Pyx_PyErr_ExceptionMatches(err)  PyErr_ExceptionMatches(err)
+#endif
+
+/* GetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_GetException(type, value, tb)  __Pyx__GetException(__pyx_tstate, type, value, tb)
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb);
+#endif
+
+/* TypeImport.proto */
+#ifndef __PYX_HAVE_RT_ImportType_proto
+#define __PYX_HAVE_RT_ImportType_proto
+enum __Pyx_ImportType_CheckSize {
+   __Pyx_ImportType_CheckSize_Error = 0,
+   __Pyx_ImportType_CheckSize_Warn = 1,
+   __Pyx_ImportType_CheckSize_Ignore = 2
+};
+static PyTypeObject *__Pyx_ImportType(PyObject* module, const char *module_name, const char *class_name, size_t size, enum __Pyx_ImportType_CheckSize check_size);
+#endif
+
+/* Import.proto */
+static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level);
+
+/* CLineInTraceback.proto */
+#ifdef CYTHON_CLINE_IN_TRACEBACK
+#define __Pyx_CLineForTraceback(tstate, c_line)  (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0)
+#else
+static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line);
+#endif
+
+/* CodeObjectCache.proto */
+typedef struct {
+    PyCodeObject* code_object;
+    int code_line;
+} __Pyx_CodeObjectCacheEntry;
+struct __Pyx_CodeObjectCache {
+    int count;
+    int max_count;
+    __Pyx_CodeObjectCacheEntry* entries;
+};
+static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL};
+static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line);
+static PyCodeObject *__pyx_find_code_object(int code_line);
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object);
+
+/* AddTraceback.proto */
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename);
+
+/* BufferStructDeclare.proto */
+typedef struct {
+  Py_ssize_t shape, strides, suboffsets;
+} __Pyx_Buf_DimInfo;
+typedef struct {
+  size_t refcount;
+  Py_buffer pybuffer;
+} __Pyx_Buffer;
+typedef struct {
+  __Pyx_Buffer *rcbuffer;
+  char *data;
+  __Pyx_Buf_DimInfo diminfo[8];
+} __Pyx_LocalBuf_ND;
+
+#if PY_MAJOR_VERSION < 3
+    static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags);
+    static void __Pyx_ReleaseBuffer(Py_buffer *view);
+#else
+    #define __Pyx_GetBuffer PyObject_GetBuffer
+    #define __Pyx_ReleaseBuffer PyBuffer_Release
+#endif
+
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value);
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value);
+
+/* RealImag.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #define __Pyx_CREAL(z) ((z).real())
+    #define __Pyx_CIMAG(z) ((z).imag())
+  #else
+    #define __Pyx_CREAL(z) (__real__(z))
+    #define __Pyx_CIMAG(z) (__imag__(z))
+  #endif
+#else
+    #define __Pyx_CREAL(z) ((z).real)
+    #define __Pyx_CIMAG(z) ((z).imag)
+#endif
+#if defined(__cplusplus) && CYTHON_CCOMPLEX\
+        && (defined(_WIN32) || defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 5 || __GNUC__ == 4 && __GNUC_MINOR__ >= 4 )) || __cplusplus >= 201103)
+    #define __Pyx_SET_CREAL(z,x) ((z).real(x))
+    #define __Pyx_SET_CIMAG(z,y) ((z).imag(y))
+#else
+    #define __Pyx_SET_CREAL(z,x) __Pyx_CREAL(z) = (x)
+    #define __Pyx_SET_CIMAG(z,y) __Pyx_CIMAG(z) = (y)
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_float(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_float(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_float(a, b) ((a)-(b))
+    #define __Pyx_c_prod_float(a, b) ((a)*(b))
+    #define __Pyx_c_quot_float(a, b) ((a)/(b))
+    #define __Pyx_c_neg_float(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_float(z) ((z)==(float)0)
+    #define __Pyx_c_conj_float(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (::std::abs(z))
+        #define __Pyx_c_pow_float(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_float(z) ((z)==0)
+    #define __Pyx_c_conj_float(z)    (conjf(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (cabsf(z))
+        #define __Pyx_c_pow_float(a, b)  (cpowf(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex);
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex);
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    #endif
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_double(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_double(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_double(a, b) ((a)-(b))
+    #define __Pyx_c_prod_double(a, b) ((a)*(b))
+    #define __Pyx_c_quot_double(a, b) ((a)/(b))
+    #define __Pyx_c_neg_double(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_double(z) ((z)==(double)0)
+    #define __Pyx_c_conj_double(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (::std::abs(z))
+        #define __Pyx_c_pow_double(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_double(z) ((z)==0)
+    #define __Pyx_c_conj_double(z)    (conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (cabs(z))
+        #define __Pyx_c_pow_double(a, b)  (cpow(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex);
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex);
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    #endif
+#endif
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_enum__NPY_TYPES(enum NPY_TYPES value);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE unsigned int __Pyx_PyInt_As_unsigned_int(PyObject *);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *);
+
+/* FastTypeChecks.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type)
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2);
+#else
+#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type)
+#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type)
+#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2))
+#endif
+#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception)
+
+/* CheckBinaryVersion.proto */
+static int __Pyx_check_binary_version(void);
+
+/* InitStrings.proto */
+static int __Pyx_InitStrings(__Pyx_StringTabEntry *t);
+
+
+/* Module declarations from 'cpython.buffer' */
+
+/* Module declarations from 'libc.string' */
+
+/* Module declarations from 'libc.stdio' */
+
+/* Module declarations from '__builtin__' */
+
+/* Module declarations from 'cpython.type' */
+static PyTypeObject *__pyx_ptype_7cpython_4type_type = 0;
+
+/* Module declarations from 'cpython' */
+
+/* Module declarations from 'cpython.object' */
+
+/* Module declarations from 'cpython.ref' */
+
+/* Module declarations from 'cpython.mem' */
+
+/* Module declarations from 'numpy' */
+
+/* Module declarations from 'numpy' */
+static PyTypeObject *__pyx_ptype_5numpy_dtype = 0;
+static PyTypeObject *__pyx_ptype_5numpy_flatiter = 0;
+static PyTypeObject *__pyx_ptype_5numpy_broadcast = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ndarray = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ufunc = 0;
+static CYTHON_INLINE char *__pyx_f_5numpy__util_dtypestring(PyArray_Descr *, char *, char *, int *); /*proto*/
+
+/* Module declarations from 'mmdet.ops.nms.soft_nms_cpu' */
+static CYTHON_INLINE __pyx_t_5numpy_float32_t __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_max(__pyx_t_5numpy_float32_t, __pyx_t_5numpy_float32_t); /*proto*/
+static CYTHON_INLINE __pyx_t_5numpy_float32_t __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_min(__pyx_t_5numpy_float32_t, __pyx_t_5numpy_float32_t); /*proto*/
+static __Pyx_TypeInfo __Pyx_TypeInfo_float = { "float", NULL, sizeof(float), { 0 }, 0, 'R', 0, 0 };
+#define __Pyx_MODULE_NAME "mmdet.ops.nms.soft_nms_cpu"
+extern int __pyx_module_is_main_mmdet__ops__nms__soft_nms_cpu;
+int __pyx_module_is_main_mmdet__ops__nms__soft_nms_cpu = 0;
+
+/* Implementation of 'mmdet.ops.nms.soft_nms_cpu' */
+static PyObject *__pyx_builtin_range;
+static PyObject *__pyx_builtin_ValueError;
+static PyObject *__pyx_builtin_RuntimeError;
+static PyObject *__pyx_builtin_ImportError;
+static const char __pyx_k_N[] = "N";
+static const char __pyx_k_i[] = "i";
+static const char __pyx_k_s[] = "s";
+static const char __pyx_k_ih[] = "ih";
+static const char __pyx_k_iw[] = "iw";
+static const char __pyx_k_np[] = "np";
+static const char __pyx_k_ov[] = "ov";
+static const char __pyx_k_ti[] = "ti";
+static const char __pyx_k_ts[] = "ts";
+static const char __pyx_k_ua[] = "ua";
+static const char __pyx_k_x1[] = "x1";
+static const char __pyx_k_x2[] = "x2";
+static const char __pyx_k_y1[] = "y1";
+static const char __pyx_k_y2[] = "y2";
+static const char __pyx_k_exp[] = "exp";
+static const char __pyx_k_pos[] = "pos";
+static const char __pyx_k_tx1[] = "tx1";
+static const char __pyx_k_tx2[] = "tx2";
+static const char __pyx_k_ty1[] = "ty1";
+static const char __pyx_k_ty2[] = "ty2";
+static const char __pyx_k_area[] = "area";
+static const char __pyx_k_copy[] = "copy";
+static const char __pyx_k_inds[] = "inds";
+static const char __pyx_k_main[] = "__main__";
+static const char __pyx_k_name[] = "__name__";
+static const char __pyx_k_test[] = "__test__";
+static const char __pyx_k_boxes[] = "boxes";
+static const char __pyx_k_numpy[] = "numpy";
+static const char __pyx_k_range[] = "range";
+static const char __pyx_k_shape[] = "shape";
+static const char __pyx_k_sigma[] = "sigma";
+static const char __pyx_k_arange[] = "arange";
+static const char __pyx_k_import[] = "__import__";
+static const char __pyx_k_maxpos[] = "maxpos";
+static const char __pyx_k_method[] = "method";
+static const char __pyx_k_weight[] = "weight";
+static const char __pyx_k_iou_thr[] = "iou_thr";
+static const char __pyx_k_box_area[] = "box_area";
+static const char __pyx_k_boxes_in[] = "boxes_in";
+static const char __pyx_k_maxscore[] = "maxscore";
+static const char __pyx_k_min_score[] = "min_score";
+static const char __pyx_k_ValueError[] = "ValueError";
+static const char __pyx_k_ImportError[] = "ImportError";
+static const char __pyx_k_RuntimeError[] = "RuntimeError";
+static const char __pyx_k_soft_nms_cpu[] = "soft_nms_cpu";
+static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback";
+static const char __pyx_k_mmdet_ops_nms_soft_nms_cpu[] = "mmdet.ops.nms.soft_nms_cpu";
+static const char __pyx_k_ndarray_is_not_C_contiguous[] = "ndarray is not C contiguous";
+static const char __pyx_k_numpy_core_multiarray_failed_to[] = "numpy.core.multiarray failed to import";
+static const char __pyx_k_unknown_dtype_code_in_numpy_pxd[] = "unknown dtype code in numpy.pxd (%d)";
+static const char __pyx_k_Format_string_allocated_too_shor[] = "Format string allocated too short, see comment in numpy.pxd";
+static const char __pyx_k_Non_native_byte_order_not_suppor[] = "Non-native byte order not supported";
+static const char __pyx_k_mmdet_ops_nms_src_soft_nms_cpu_p[] = "mmdet/ops/nms/src/soft_nms_cpu.pyx";
+static const char __pyx_k_ndarray_is_not_Fortran_contiguou[] = "ndarray is not Fortran contiguous";
+static const char __pyx_k_numpy_core_umath_failed_to_impor[] = "numpy.core.umath failed to import";
+static const char __pyx_k_Format_string_allocated_too_shor_2[] = "Format string allocated too short.";
+static PyObject *__pyx_kp_u_Format_string_allocated_too_shor;
+static PyObject *__pyx_kp_u_Format_string_allocated_too_shor_2;
+static PyObject *__pyx_n_s_ImportError;
+static PyObject *__pyx_n_s_N;
+static PyObject *__pyx_kp_u_Non_native_byte_order_not_suppor;
+static PyObject *__pyx_n_s_RuntimeError;
+static PyObject *__pyx_n_s_ValueError;
+static PyObject *__pyx_n_s_arange;
+static PyObject *__pyx_n_s_area;
+static PyObject *__pyx_n_s_box_area;
+static PyObject *__pyx_n_s_boxes;
+static PyObject *__pyx_n_s_boxes_in;
+static PyObject *__pyx_n_s_cline_in_traceback;
+static PyObject *__pyx_n_s_copy;
+static PyObject *__pyx_n_s_exp;
+static PyObject *__pyx_n_s_i;
+static PyObject *__pyx_n_s_ih;
+static PyObject *__pyx_n_s_import;
+static PyObject *__pyx_n_s_inds;
+static PyObject *__pyx_n_s_iou_thr;
+static PyObject *__pyx_n_s_iw;
+static PyObject *__pyx_n_s_main;
+static PyObject *__pyx_n_s_maxpos;
+static PyObject *__pyx_n_s_maxscore;
+static PyObject *__pyx_n_s_method;
+static PyObject *__pyx_n_s_min_score;
+static PyObject *__pyx_n_s_mmdet_ops_nms_soft_nms_cpu;
+static PyObject *__pyx_kp_s_mmdet_ops_nms_src_soft_nms_cpu_p;
+static PyObject *__pyx_n_s_name;
+static PyObject *__pyx_kp_u_ndarray_is_not_C_contiguous;
+static PyObject *__pyx_kp_u_ndarray_is_not_Fortran_contiguou;
+static PyObject *__pyx_n_s_np;
+static PyObject *__pyx_n_s_numpy;
+static PyObject *__pyx_kp_u_numpy_core_multiarray_failed_to;
+static PyObject *__pyx_kp_u_numpy_core_umath_failed_to_impor;
+static PyObject *__pyx_n_s_ov;
+static PyObject *__pyx_n_s_pos;
+static PyObject *__pyx_n_s_range;
+static PyObject *__pyx_n_s_s;
+static PyObject *__pyx_n_s_shape;
+static PyObject *__pyx_n_s_sigma;
+static PyObject *__pyx_n_s_soft_nms_cpu;
+static PyObject *__pyx_n_s_test;
+static PyObject *__pyx_n_s_ti;
+static PyObject *__pyx_n_s_ts;
+static PyObject *__pyx_n_s_tx1;
+static PyObject *__pyx_n_s_tx2;
+static PyObject *__pyx_n_s_ty1;
+static PyObject *__pyx_n_s_ty2;
+static PyObject *__pyx_n_s_ua;
+static PyObject *__pyx_kp_u_unknown_dtype_code_in_numpy_pxd;
+static PyObject *__pyx_n_s_weight;
+static PyObject *__pyx_n_s_x1;
+static PyObject *__pyx_n_s_x2;
+static PyObject *__pyx_n_s_y1;
+static PyObject *__pyx_n_s_y2;
+static PyObject *__pyx_pf_5mmdet_3ops_3nms_12soft_nms_cpu_soft_nms_cpu(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_boxes_in, float __pyx_v_iou_thr, unsigned int __pyx_v_method, float __pyx_v_sigma, float __pyx_v_min_score); /* proto */
+static int __pyx_pf_5numpy_7ndarray___getbuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */
+static void __pyx_pf_5numpy_7ndarray_2__releasebuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info); /* proto */
+static PyObject *__pyx_int_0;
+static PyObject *__pyx_int_1;
+static PyObject *__pyx_int_2;
+static PyObject *__pyx_int_3;
+static PyObject *__pyx_int_4;
+static PyObject *__pyx_tuple_;
+static PyObject *__pyx_tuple__2;
+static PyObject *__pyx_tuple__3;
+static PyObject *__pyx_tuple__4;
+static PyObject *__pyx_tuple__5;
+static PyObject *__pyx_tuple__6;
+static PyObject *__pyx_tuple__7;
+static PyObject *__pyx_tuple__8;
+static PyObject *__pyx_codeobj__9;
+/* Late includes */
+
+/* "mmdet/ops/nms/src/soft_nms_cpu.pyx":15
+ * 
+ * 
+ * cdef inline np.float32_t max(np.float32_t a, np.float32_t b):             # <<<<<<<<<<<<<<
+ *     return a if a >= b else b
+ * 
+ */
+
+static CYTHON_INLINE __pyx_t_5numpy_float32_t __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_max(__pyx_t_5numpy_float32_t __pyx_v_a, __pyx_t_5numpy_float32_t __pyx_v_b) {
+  __pyx_t_5numpy_float32_t __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __pyx_t_5numpy_float32_t __pyx_t_1;
+  __Pyx_RefNannySetupContext("max", 0);
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":16
+ * 
+ * cdef inline np.float32_t max(np.float32_t a, np.float32_t b):
+ *     return a if a >= b else b             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline np.float32_t min(np.float32_t a, np.float32_t b):
+ */
+  if (((__pyx_v_a >= __pyx_v_b) != 0)) {
+    __pyx_t_1 = __pyx_v_a;
+  } else {
+    __pyx_t_1 = __pyx_v_b;
+  }
+  __pyx_r = __pyx_t_1;
+  goto __pyx_L0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":15
+ * 
+ * 
+ * cdef inline np.float32_t max(np.float32_t a, np.float32_t b):             # <<<<<<<<<<<<<<
+ *     return a if a >= b else b
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mmdet/ops/nms/src/soft_nms_cpu.pyx":18
+ *     return a if a >= b else b
+ * 
+ * cdef inline np.float32_t min(np.float32_t a, np.float32_t b):             # <<<<<<<<<<<<<<
+ *     return a if a <= b else b
+ * 
+ */
+
+static CYTHON_INLINE __pyx_t_5numpy_float32_t __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_min(__pyx_t_5numpy_float32_t __pyx_v_a, __pyx_t_5numpy_float32_t __pyx_v_b) {
+  __pyx_t_5numpy_float32_t __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __pyx_t_5numpy_float32_t __pyx_t_1;
+  __Pyx_RefNannySetupContext("min", 0);
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":19
+ * 
+ * cdef inline np.float32_t min(np.float32_t a, np.float32_t b):
+ *     return a if a <= b else b             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  if (((__pyx_v_a <= __pyx_v_b) != 0)) {
+    __pyx_t_1 = __pyx_v_a;
+  } else {
+    __pyx_t_1 = __pyx_v_b;
+  }
+  __pyx_r = __pyx_t_1;
+  goto __pyx_L0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":18
+ *     return a if a >= b else b
+ * 
+ * cdef inline np.float32_t min(np.float32_t a, np.float32_t b):             # <<<<<<<<<<<<<<
+ *     return a if a <= b else b
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mmdet/ops/nms/src/soft_nms_cpu.pyx":22
+ * 
+ * 
+ * def soft_nms_cpu(             # <<<<<<<<<<<<<<
+ *     np.ndarray[float, ndim=2] boxes_in,
+ *     float iou_thr,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_5mmdet_3ops_3nms_12soft_nms_cpu_1soft_nms_cpu(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_5mmdet_3ops_3nms_12soft_nms_cpu_1soft_nms_cpu = {"soft_nms_cpu", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_5mmdet_3ops_3nms_12soft_nms_cpu_1soft_nms_cpu, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_5mmdet_3ops_3nms_12soft_nms_cpu_1soft_nms_cpu(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_boxes_in = 0;
+  float __pyx_v_iou_thr;
+  unsigned int __pyx_v_method;
+  float __pyx_v_sigma;
+  float __pyx_v_min_score;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("soft_nms_cpu (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_boxes_in,&__pyx_n_s_iou_thr,&__pyx_n_s_method,&__pyx_n_s_sigma,&__pyx_n_s_min_score,0};
+    PyObject* values[5] = {0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_boxes_in)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_iou_thr)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("soft_nms_cpu", 0, 2, 5, 1); __PYX_ERR(0, 22, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (kw_args > 0) {
+          PyObject* value = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_method);
+          if (value) { values[2] = value; kw_args--; }
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (kw_args > 0) {
+          PyObject* value = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_sigma);
+          if (value) { values[3] = value; kw_args--; }
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (kw_args > 0) {
+          PyObject* value = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_min_score);
+          if (value) { values[4] = value; kw_args--; }
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "soft_nms_cpu") < 0)) __PYX_ERR(0, 22, __pyx_L3_error)
+      }
+    } else {
+      switch (PyTuple_GET_SIZE(__pyx_args)) {
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+    }
+    __pyx_v_boxes_in = ((PyArrayObject *)values[0]);
+    __pyx_v_iou_thr = __pyx_PyFloat_AsFloat(values[1]); if (unlikely((__pyx_v_iou_thr == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 24, __pyx_L3_error)
+    if (values[2]) {
+      __pyx_v_method = __Pyx_PyInt_As_unsigned_int(values[2]); if (unlikely((__pyx_v_method == (unsigned int)-1) && PyErr_Occurred())) __PYX_ERR(0, 25, __pyx_L3_error)
+    } else {
+      __pyx_v_method = ((unsigned int)1);
+    }
+    if (values[3]) {
+      __pyx_v_sigma = __pyx_PyFloat_AsFloat(values[3]); if (unlikely((__pyx_v_sigma == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 26, __pyx_L3_error)
+    } else {
+      __pyx_v_sigma = ((float)0.5);
+    }
+    if (values[4]) {
+      __pyx_v_min_score = __pyx_PyFloat_AsFloat(values[4]); if (unlikely((__pyx_v_min_score == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error)
+    } else {
+      __pyx_v_min_score = ((float)0.001);
+    }
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("soft_nms_cpu", 0, 2, 5, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 22, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mmdet.ops.nms.soft_nms_cpu.soft_nms_cpu", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_boxes_in), __pyx_ptype_5numpy_ndarray, 1, "boxes_in", 0))) __PYX_ERR(0, 23, __pyx_L1_error)
+  __pyx_r = __pyx_pf_5mmdet_3ops_3nms_12soft_nms_cpu_soft_nms_cpu(__pyx_self, __pyx_v_boxes_in, __pyx_v_iou_thr, __pyx_v_method, __pyx_v_sigma, __pyx_v_min_score);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_5mmdet_3ops_3nms_12soft_nms_cpu_soft_nms_cpu(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_boxes_in, float __pyx_v_iou_thr, unsigned int __pyx_v_method, float __pyx_v_sigma, float __pyx_v_min_score) {
+  PyObject *__pyx_v_boxes = NULL;
+  int __pyx_v_N;
+  float __pyx_v_iw;
+  float __pyx_v_ih;
+  float __pyx_v_ua;
+  int __pyx_v_pos;
+  float __pyx_v_maxscore;
+  int __pyx_v_maxpos;
+  float __pyx_v_x1;
+  float __pyx_v_x2;
+  float __pyx_v_y1;
+  float __pyx_v_y2;
+  float __pyx_v_tx1;
+  float __pyx_v_tx2;
+  float __pyx_v_ty1;
+  float __pyx_v_ty2;
+  float __pyx_v_ts;
+  float __pyx_v_area;
+  float __pyx_v_weight;
+  float __pyx_v_ov;
+  PyObject *__pyx_v_inds = NULL;
+  PyObject *__pyx_v_i = NULL;
+  PyObject *__pyx_v_ti = NULL;
+  CYTHON_UNUSED PyObject *__pyx_v_s = NULL;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_boxes_in;
+  __Pyx_Buffer __pyx_pybuffer_boxes_in;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  Py_ssize_t __pyx_t_6;
+  PyObject *(*__pyx_t_7)(PyObject *);
+  float __pyx_t_8;
+  int __pyx_t_9;
+  PyObject *__pyx_t_10 = NULL;
+  long __pyx_t_11;
+  __Pyx_RefNannySetupContext("soft_nms_cpu", 0);
+  __pyx_pybuffer_boxes_in.pybuffer.buf = NULL;
+  __pyx_pybuffer_boxes_in.refcount = 0;
+  __pyx_pybuffernd_boxes_in.data = NULL;
+  __pyx_pybuffernd_boxes_in.rcbuffer = &__pyx_pybuffer_boxes_in;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_boxes_in.rcbuffer->pybuffer, (PyObject*)__pyx_v_boxes_in, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 22, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_boxes_in.diminfo[0].strides = __pyx_pybuffernd_boxes_in.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_boxes_in.diminfo[0].shape = __pyx_pybuffernd_boxes_in.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_boxes_in.diminfo[1].strides = __pyx_pybuffernd_boxes_in.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_boxes_in.diminfo[1].shape = __pyx_pybuffernd_boxes_in.rcbuffer->pybuffer.shape[1];
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":29
+ *     float min_score=0.001,
+ * ):
+ *     boxes = boxes_in.copy()             # <<<<<<<<<<<<<<
+ *     cdef int N = boxes.shape[0]
+ *     cdef float iw, ih, box_area
+ */
+  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_boxes_in), __pyx_n_s_copy); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 29, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __pyx_t_3 = NULL;
+  if (CYTHON_UNPACK_METHODS && likely(PyMethod_Check(__pyx_t_2))) {
+    __pyx_t_3 = PyMethod_GET_SELF(__pyx_t_2);
+    if (likely(__pyx_t_3)) {
+      PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_2);
+      __Pyx_INCREF(__pyx_t_3);
+      __Pyx_INCREF(function);
+      __Pyx_DECREF_SET(__pyx_t_2, function);
+    }
+  }
+  __pyx_t_1 = (__pyx_t_3) ? __Pyx_PyObject_CallOneArg(__pyx_t_2, __pyx_t_3) : __Pyx_PyObject_CallNoArg(__pyx_t_2);
+  __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+  if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 29, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __pyx_v_boxes = __pyx_t_1;
+  __pyx_t_1 = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":30
+ * ):
+ *     boxes = boxes_in.copy()
+ *     cdef int N = boxes.shape[0]             # <<<<<<<<<<<<<<
+ *     cdef float iw, ih, box_area
+ *     cdef float ua
+ */
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_v_boxes, __pyx_n_s_shape); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 30, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_t_2 = __Pyx_GetItemInt(__pyx_t_1, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 30, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_2); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 30, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __pyx_v_N = __pyx_t_4;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":33
+ *     cdef float iw, ih, box_area
+ *     cdef float ua
+ *     cdef int pos = 0             # <<<<<<<<<<<<<<
+ *     cdef float maxscore = 0
+ *     cdef int maxpos = 0
+ */
+  __pyx_v_pos = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":34
+ *     cdef float ua
+ *     cdef int pos = 0
+ *     cdef float maxscore = 0             # <<<<<<<<<<<<<<
+ *     cdef int maxpos = 0
+ *     cdef float x1, x2, y1, y2, tx1, tx2, ty1, ty2, ts, area, weight, ov
+ */
+  __pyx_v_maxscore = 0.0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":35
+ *     cdef int pos = 0
+ *     cdef float maxscore = 0
+ *     cdef int maxpos = 0             # <<<<<<<<<<<<<<
+ *     cdef float x1, x2, y1, y2, tx1, tx2, ty1, ty2, ts, area, weight, ov
+ *     inds = np.arange(N)
+ */
+  __pyx_v_maxpos = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":37
+ *     cdef int maxpos = 0
+ *     cdef float x1, x2, y1, y2, tx1, tx2, ty1, ty2, ts, area, weight, ov
+ *     inds = np.arange(N)             # <<<<<<<<<<<<<<
+ * 
+ *     for i in range(N):
+ */
+  __Pyx_GetModuleGlobalName(__pyx_t_1, __pyx_n_s_np); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 37, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_arange); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 37, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_3);
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_N); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 37, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_t_5 = NULL;
+  if (CYTHON_UNPACK_METHODS && unlikely(PyMethod_Check(__pyx_t_3))) {
+    __pyx_t_5 = PyMethod_GET_SELF(__pyx_t_3);
+    if (likely(__pyx_t_5)) {
+      PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3);
+      __Pyx_INCREF(__pyx_t_5);
+      __Pyx_INCREF(function);
+      __Pyx_DECREF_SET(__pyx_t_3, function);
+    }
+  }
+  __pyx_t_2 = (__pyx_t_5) ? __Pyx_PyObject_Call2Args(__pyx_t_3, __pyx_t_5, __pyx_t_1) : __Pyx_PyObject_CallOneArg(__pyx_t_3, __pyx_t_1);
+  __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0;
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 37, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+  __pyx_v_inds = __pyx_t_2;
+  __pyx_t_2 = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":39
+ *     inds = np.arange(N)
+ * 
+ *     for i in range(N):             # <<<<<<<<<<<<<<
+ *         maxscore = boxes[i, 4]
+ *         maxpos = i
+ */
+  __pyx_t_2 = __Pyx_PyInt_From_int(__pyx_v_N); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 39, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_range, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 39, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_3);
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  if (likely(PyList_CheckExact(__pyx_t_3)) || PyTuple_CheckExact(__pyx_t_3)) {
+    __pyx_t_2 = __pyx_t_3; __Pyx_INCREF(__pyx_t_2); __pyx_t_6 = 0;
+    __pyx_t_7 = NULL;
+  } else {
+    __pyx_t_6 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_t_3); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 39, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_2);
+    __pyx_t_7 = Py_TYPE(__pyx_t_2)->tp_iternext; if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 39, __pyx_L1_error)
+  }
+  __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+  for (;;) {
+    if (likely(!__pyx_t_7)) {
+      if (likely(PyList_CheckExact(__pyx_t_2))) {
+        if (__pyx_t_6 >= PyList_GET_SIZE(__pyx_t_2)) break;
+        #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+        __pyx_t_3 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_3); __pyx_t_6++; if (unlikely(0 < 0)) __PYX_ERR(0, 39, __pyx_L1_error)
+        #else
+        __pyx_t_3 = PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 39, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_3);
+        #endif
+      } else {
+        if (__pyx_t_6 >= PyTuple_GET_SIZE(__pyx_t_2)) break;
+        #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+        __pyx_t_3 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_3); __pyx_t_6++; if (unlikely(0 < 0)) __PYX_ERR(0, 39, __pyx_L1_error)
+        #else
+        __pyx_t_3 = PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 39, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_3);
+        #endif
+      }
+    } else {
+      __pyx_t_3 = __pyx_t_7(__pyx_t_2);
+      if (unlikely(!__pyx_t_3)) {
+        PyObject* exc_type = PyErr_Occurred();
+        if (exc_type) {
+          if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear();
+          else __PYX_ERR(0, 39, __pyx_L1_error)
+        }
+        break;
+      }
+      __Pyx_GOTREF(__pyx_t_3);
+    }
+    __Pyx_XDECREF_SET(__pyx_v_i, __pyx_t_3);
+    __pyx_t_3 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":40
+ * 
+ *     for i in range(N):
+ *         maxscore = boxes[i, 4]             # <<<<<<<<<<<<<<
+ *         maxpos = i
+ * 
+ */
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 40, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_4);
+    __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 40, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_1); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 40, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_v_maxscore = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":41
+ *     for i in range(N):
+ *         maxscore = boxes[i, 4]
+ *         maxpos = i             # <<<<<<<<<<<<<<
+ * 
+ *         tx1 = boxes[i, 0]
+ */
+    __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_v_i); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 41, __pyx_L1_error)
+    __pyx_v_maxpos = __pyx_t_4;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":43
+ *         maxpos = i
+ * 
+ *         tx1 = boxes[i, 0]             # <<<<<<<<<<<<<<
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]
+ */
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 43, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_0);
+    __Pyx_GIVEREF(__pyx_int_0);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_0);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 43, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 43, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_tx1 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":44
+ * 
+ *         tx1 = boxes[i, 0]
+ *         ty1 = boxes[i, 1]             # <<<<<<<<<<<<<<
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]
+ */
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 44, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_1);
+    __Pyx_GIVEREF(__pyx_int_1);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_1);
+    __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 44, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_1); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 44, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_v_ty1 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":45
+ *         tx1 = boxes[i, 0]
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]             # <<<<<<<<<<<<<<
+ *         ty2 = boxes[i, 3]
+ *         ts = boxes[i, 4]
+ */
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 45, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_2);
+    __Pyx_GIVEREF(__pyx_int_2);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_2);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 45, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 45, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_tx2 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":46
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]             # <<<<<<<<<<<<<<
+ *         ts = boxes[i, 4]
+ *         ti = inds[i]
+ */
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 46, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_3);
+    __Pyx_GIVEREF(__pyx_int_3);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_3);
+    __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 46, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_1); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 46, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_v_ty2 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":47
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]
+ *         ts = boxes[i, 4]             # <<<<<<<<<<<<<<
+ *         ti = inds[i]
+ * 
+ */
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 47, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_4);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 47, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 47, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_ts = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":48
+ *         ty2 = boxes[i, 3]
+ *         ts = boxes[i, 4]
+ *         ti = inds[i]             # <<<<<<<<<<<<<<
+ * 
+ *         pos = i + 1
+ */
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_inds, __pyx_v_i); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 48, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_XDECREF_SET(__pyx_v_ti, __pyx_t_3);
+    __pyx_t_3 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":50
+ *         ti = inds[i]
+ * 
+ *         pos = i + 1             # <<<<<<<<<<<<<<
+ *         # get max box
+ *         while pos < N:
+ */
+    __pyx_t_3 = __Pyx_PyInt_AddObjC(__pyx_v_i, __pyx_int_1, 1, 0, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 50, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_pos = __pyx_t_4;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":52
+ *         pos = i + 1
+ *         # get max box
+ *         while pos < N:             # <<<<<<<<<<<<<<
+ *             if maxscore < boxes[pos, 4]:
+ *                 maxscore = boxes[pos, 4]
+ */
+    while (1) {
+      __pyx_t_9 = ((__pyx_v_pos < __pyx_v_N) != 0);
+      if (!__pyx_t_9) break;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":53
+ *         # get max box
+ *         while pos < N:
+ *             if maxscore < boxes[pos, 4]:             # <<<<<<<<<<<<<<
+ *                 maxscore = boxes[pos, 4]
+ *                 maxpos = pos
+ */
+      __pyx_t_3 = PyFloat_FromDouble(__pyx_v_maxscore); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_1);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_1);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_1);
+      __Pyx_INCREF(__pyx_int_4);
+      __Pyx_GIVEREF(__pyx_int_4);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_4);
+      __pyx_t_1 = 0;
+      __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_1);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __pyx_t_5 = PyObject_RichCompare(__pyx_t_3, __pyx_t_1, Py_LT); __Pyx_XGOTREF(__pyx_t_5); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+      __pyx_t_9 = __Pyx_PyObject_IsTrue(__pyx_t_5); if (unlikely(__pyx_t_9 < 0)) __PYX_ERR(0, 53, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      if (__pyx_t_9) {
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":54
+ *         while pos < N:
+ *             if maxscore < boxes[pos, 4]:
+ *                 maxscore = boxes[pos, 4]             # <<<<<<<<<<<<<<
+ *                 maxpos = pos
+ *             pos = pos + 1
+ */
+        __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 54, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_5);
+        __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 54, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_1);
+        __Pyx_GIVEREF(__pyx_t_5);
+        PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+        __Pyx_INCREF(__pyx_int_4);
+        __Pyx_GIVEREF(__pyx_int_4);
+        PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_4);
+        __pyx_t_5 = 0;
+        __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 54, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_5);
+        __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+        __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_5); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 54, __pyx_L1_error)
+        __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+        __pyx_v_maxscore = __pyx_t_8;
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":55
+ *             if maxscore < boxes[pos, 4]:
+ *                 maxscore = boxes[pos, 4]
+ *                 maxpos = pos             # <<<<<<<<<<<<<<
+ *             pos = pos + 1
+ * 
+ */
+        __pyx_v_maxpos = __pyx_v_pos;
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":53
+ *         # get max box
+ *         while pos < N:
+ *             if maxscore < boxes[pos, 4]:             # <<<<<<<<<<<<<<
+ *                 maxscore = boxes[pos, 4]
+ *                 maxpos = pos
+ */
+      }
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":56
+ *                 maxscore = boxes[pos, 4]
+ *                 maxpos = pos
+ *             pos = pos + 1             # <<<<<<<<<<<<<<
+ * 
+ *         # add max box as a detection
+ */
+      __pyx_v_pos = (__pyx_v_pos + 1);
+    }
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":59
+ * 
+ *         # add max box as a detection
+ *         boxes[i, 0] = boxes[maxpos, 0]             # <<<<<<<<<<<<<<
+ *         boxes[i, 1] = boxes[maxpos, 1]
+ *         boxes[i, 2] = boxes[maxpos, 2]
+ */
+    __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 59, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 59, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_5);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+    __Pyx_INCREF(__pyx_int_0);
+    __Pyx_GIVEREF(__pyx_int_0);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_0);
+    __pyx_t_5 = 0;
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 59, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 59, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_0);
+    __Pyx_GIVEREF(__pyx_int_0);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_0);
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 59, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":60
+ *         # add max box as a detection
+ *         boxes[i, 0] = boxes[maxpos, 0]
+ *         boxes[i, 1] = boxes[maxpos, 1]             # <<<<<<<<<<<<<<
+ *         boxes[i, 2] = boxes[maxpos, 2]
+ *         boxes[i, 3] = boxes[maxpos, 3]
+ */
+    __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 60, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 60, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_5);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+    __Pyx_INCREF(__pyx_int_1);
+    __Pyx_GIVEREF(__pyx_int_1);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_1);
+    __pyx_t_5 = 0;
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 60, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 60, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_1);
+    __Pyx_GIVEREF(__pyx_int_1);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_1);
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 60, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":61
+ *         boxes[i, 0] = boxes[maxpos, 0]
+ *         boxes[i, 1] = boxes[maxpos, 1]
+ *         boxes[i, 2] = boxes[maxpos, 2]             # <<<<<<<<<<<<<<
+ *         boxes[i, 3] = boxes[maxpos, 3]
+ *         boxes[i, 4] = boxes[maxpos, 4]
+ */
+    __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 61, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 61, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_5);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+    __Pyx_INCREF(__pyx_int_2);
+    __Pyx_GIVEREF(__pyx_int_2);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_2);
+    __pyx_t_5 = 0;
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 61, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 61, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_2);
+    __Pyx_GIVEREF(__pyx_int_2);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_2);
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 61, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":62
+ *         boxes[i, 1] = boxes[maxpos, 1]
+ *         boxes[i, 2] = boxes[maxpos, 2]
+ *         boxes[i, 3] = boxes[maxpos, 3]             # <<<<<<<<<<<<<<
+ *         boxes[i, 4] = boxes[maxpos, 4]
+ *         inds[i] = inds[maxpos]
+ */
+    __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 62, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 62, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_5);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+    __Pyx_INCREF(__pyx_int_3);
+    __Pyx_GIVEREF(__pyx_int_3);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_3);
+    __pyx_t_5 = 0;
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 62, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 62, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_3);
+    __Pyx_GIVEREF(__pyx_int_3);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_3);
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 62, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":63
+ *         boxes[i, 2] = boxes[maxpos, 2]
+ *         boxes[i, 3] = boxes[maxpos, 3]
+ *         boxes[i, 4] = boxes[maxpos, 4]             # <<<<<<<<<<<<<<
+ *         inds[i] = inds[maxpos]
+ * 
+ */
+    __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 63, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 63, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_5);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_5);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_4);
+    __pyx_t_5 = 0;
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 63, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 63, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_4);
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 63, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":64
+ *         boxes[i, 3] = boxes[maxpos, 3]
+ *         boxes[i, 4] = boxes[maxpos, 4]
+ *         inds[i] = inds[maxpos]             # <<<<<<<<<<<<<<
+ * 
+ *         # swap ith box with position of max box
+ */
+    __pyx_t_5 = __Pyx_GetItemInt(__pyx_v_inds, __pyx_v_maxpos, int, 1, __Pyx_PyInt_From_int, 0, 1, 0); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 64, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    if (unlikely(PyObject_SetItem(__pyx_v_inds, __pyx_v_i, __pyx_t_5) < 0)) __PYX_ERR(0, 64, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":67
+ * 
+ *         # swap ith box with position of max box
+ *         boxes[maxpos, 0] = tx1             # <<<<<<<<<<<<<<
+ *         boxes[maxpos, 1] = ty1
+ *         boxes[maxpos, 2] = tx2
+ */
+    __pyx_t_5 = PyFloat_FromDouble(__pyx_v_tx1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 67, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 67, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 67, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_GIVEREF(__pyx_t_1);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+    __Pyx_INCREF(__pyx_int_0);
+    __Pyx_GIVEREF(__pyx_int_0);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_0);
+    __pyx_t_1 = 0;
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_5) < 0)) __PYX_ERR(0, 67, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":68
+ *         # swap ith box with position of max box
+ *         boxes[maxpos, 0] = tx1
+ *         boxes[maxpos, 1] = ty1             # <<<<<<<<<<<<<<
+ *         boxes[maxpos, 2] = tx2
+ *         boxes[maxpos, 3] = ty2
+ */
+    __pyx_t_5 = PyFloat_FromDouble(__pyx_v_ty1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 68, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 68, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 68, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_3);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_3);
+    __Pyx_INCREF(__pyx_int_1);
+    __Pyx_GIVEREF(__pyx_int_1);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_1);
+    __pyx_t_3 = 0;
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 68, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":69
+ *         boxes[maxpos, 0] = tx1
+ *         boxes[maxpos, 1] = ty1
+ *         boxes[maxpos, 2] = tx2             # <<<<<<<<<<<<<<
+ *         boxes[maxpos, 3] = ty2
+ *         boxes[maxpos, 4] = ts
+ */
+    __pyx_t_5 = PyFloat_FromDouble(__pyx_v_tx2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 69, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 69, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 69, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_GIVEREF(__pyx_t_1);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+    __Pyx_INCREF(__pyx_int_2);
+    __Pyx_GIVEREF(__pyx_int_2);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_2);
+    __pyx_t_1 = 0;
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_5) < 0)) __PYX_ERR(0, 69, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":70
+ *         boxes[maxpos, 1] = ty1
+ *         boxes[maxpos, 2] = tx2
+ *         boxes[maxpos, 3] = ty2             # <<<<<<<<<<<<<<
+ *         boxes[maxpos, 4] = ts
+ *         inds[maxpos] = ti
+ */
+    __pyx_t_5 = PyFloat_FromDouble(__pyx_v_ty2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 70, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 70, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __pyx_t_1 = PyTuple_New(2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 70, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __Pyx_GIVEREF(__pyx_t_3);
+    PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_3);
+    __Pyx_INCREF(__pyx_int_3);
+    __Pyx_GIVEREF(__pyx_int_3);
+    PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_3);
+    __pyx_t_3 = 0;
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_1, __pyx_t_5) < 0)) __PYX_ERR(0, 70, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":71
+ *         boxes[maxpos, 2] = tx2
+ *         boxes[maxpos, 3] = ty2
+ *         boxes[maxpos, 4] = ts             # <<<<<<<<<<<<<<
+ *         inds[maxpos] = ti
+ * 
+ */
+    __pyx_t_5 = PyFloat_FromDouble(__pyx_v_ts); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 71, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_maxpos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 71, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_1);
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 71, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_GIVEREF(__pyx_t_1);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_4);
+    __pyx_t_1 = 0;
+    if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_5) < 0)) __PYX_ERR(0, 71, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":72
+ *         boxes[maxpos, 3] = ty2
+ *         boxes[maxpos, 4] = ts
+ *         inds[maxpos] = ti             # <<<<<<<<<<<<<<
+ * 
+ *         tx1 = boxes[i, 0]
+ */
+    if (unlikely(__Pyx_SetItemInt(__pyx_v_inds, __pyx_v_maxpos, __pyx_v_ti, int, 1, __Pyx_PyInt_From_int, 0, 1, 0) < 0)) __PYX_ERR(0, 72, __pyx_L1_error)
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":74
+ *         inds[maxpos] = ti
+ * 
+ *         tx1 = boxes[i, 0]             # <<<<<<<<<<<<<<
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]
+ */
+    __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 74, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_0);
+    __Pyx_GIVEREF(__pyx_int_0);
+    PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_0);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 74, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 74, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_tx1 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":75
+ * 
+ *         tx1 = boxes[i, 0]
+ *         ty1 = boxes[i, 1]             # <<<<<<<<<<<<<<
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]
+ */
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 75, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_1);
+    __Pyx_GIVEREF(__pyx_int_1);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_1);
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 75, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_5); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 75, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+    __pyx_v_ty1 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":76
+ *         tx1 = boxes[i, 0]
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]             # <<<<<<<<<<<<<<
+ *         ty2 = boxes[i, 3]
+ *         ts = boxes[i, 4]
+ */
+    __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 76, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_2);
+    __Pyx_GIVEREF(__pyx_int_2);
+    PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_2);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 76, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 76, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_tx2 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":77
+ *         ty1 = boxes[i, 1]
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]             # <<<<<<<<<<<<<<
+ *         ts = boxes[i, 4]
+ * 
+ */
+    __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 77, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_3);
+    __Pyx_GIVEREF(__pyx_int_3);
+    PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_3);
+    __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 77, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_5); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 77, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+    __pyx_v_ty2 = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":78
+ *         tx2 = boxes[i, 2]
+ *         ty2 = boxes[i, 3]
+ *         ts = boxes[i, 4]             # <<<<<<<<<<<<<<
+ * 
+ *         pos = i + 1
+ */
+    __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 78, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_5);
+    __Pyx_INCREF(__pyx_v_i);
+    __Pyx_GIVEREF(__pyx_v_i);
+    PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_v_i);
+    __Pyx_INCREF(__pyx_int_4);
+    __Pyx_GIVEREF(__pyx_int_4);
+    PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_4);
+    __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 78, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+    __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 78, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_ts = __pyx_t_8;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":80
+ *         ts = boxes[i, 4]
+ * 
+ *         pos = i + 1             # <<<<<<<<<<<<<<
+ *         # NMS iterations, note that N changes if detection boxes fall below
+ *         # threshold
+ */
+    __pyx_t_3 = __Pyx_PyInt_AddObjC(__pyx_v_i, __pyx_int_1, 1, 0, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 80, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 80, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_v_pos = __pyx_t_4;
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":83
+ *         # NMS iterations, note that N changes if detection boxes fall below
+ *         # threshold
+ *         while pos < N:             # <<<<<<<<<<<<<<
+ *             x1 = boxes[pos, 0]
+ *             y1 = boxes[pos, 1]
+ */
+    while (1) {
+      __pyx_t_9 = ((__pyx_v_pos < __pyx_v_N) != 0);
+      if (!__pyx_t_9) break;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":84
+ *         # threshold
+ *         while pos < N:
+ *             x1 = boxes[pos, 0]             # <<<<<<<<<<<<<<
+ *             y1 = boxes[pos, 1]
+ *             x2 = boxes[pos, 2]
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 84, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 84, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+      __Pyx_INCREF(__pyx_int_0);
+      __Pyx_GIVEREF(__pyx_int_0);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_0);
+      __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 84, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 84, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_v_x1 = __pyx_t_8;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":85
+ *         while pos < N:
+ *             x1 = boxes[pos, 0]
+ *             y1 = boxes[pos, 1]             # <<<<<<<<<<<<<<
+ *             x2 = boxes[pos, 2]
+ *             y2 = boxes[pos, 3]
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 85, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 85, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+      __Pyx_INCREF(__pyx_int_1);
+      __Pyx_GIVEREF(__pyx_int_1);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_1);
+      __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 85, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 85, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_v_y1 = __pyx_t_8;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":86
+ *             x1 = boxes[pos, 0]
+ *             y1 = boxes[pos, 1]
+ *             x2 = boxes[pos, 2]             # <<<<<<<<<<<<<<
+ *             y2 = boxes[pos, 3]
+ *             s = boxes[pos, 4]
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 86, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 86, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+      __Pyx_INCREF(__pyx_int_2);
+      __Pyx_GIVEREF(__pyx_int_2);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_2);
+      __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 86, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_v_x2 = __pyx_t_8;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":87
+ *             y1 = boxes[pos, 1]
+ *             x2 = boxes[pos, 2]
+ *             y2 = boxes[pos, 3]             # <<<<<<<<<<<<<<
+ *             s = boxes[pos, 4]
+ * 
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 87, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 87, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+      __Pyx_INCREF(__pyx_int_3);
+      __Pyx_GIVEREF(__pyx_int_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_3);
+      __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 87, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_v_y2 = __pyx_t_8;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":88
+ *             x2 = boxes[pos, 2]
+ *             y2 = boxes[pos, 3]
+ *             s = boxes[pos, 4]             # <<<<<<<<<<<<<<
+ * 
+ *             area = (x2 - x1 + 1) * (y2 - y1 + 1)
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 88, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 88, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GIVEREF(__pyx_t_3);
+      PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+      __Pyx_INCREF(__pyx_int_4);
+      __Pyx_GIVEREF(__pyx_int_4);
+      PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_4);
+      __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 88, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+      __Pyx_XDECREF_SET(__pyx_v_s, __pyx_t_3);
+      __pyx_t_3 = 0;
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":90
+ *             s = boxes[pos, 4]
+ * 
+ *             area = (x2 - x1 + 1) * (y2 - y1 + 1)             # <<<<<<<<<<<<<<
+ *             iw = (min(tx2, x2) - max(tx1, x1) + 1)
+ *             if iw > 0:
+ */
+      __pyx_v_area = (((__pyx_v_x2 - __pyx_v_x1) + 1.0) * ((__pyx_v_y2 - __pyx_v_y1) + 1.0));
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":91
+ * 
+ *             area = (x2 - x1 + 1) * (y2 - y1 + 1)
+ *             iw = (min(tx2, x2) - max(tx1, x1) + 1)             # <<<<<<<<<<<<<<
+ *             if iw > 0:
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ */
+      __pyx_v_iw = ((__pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_min(__pyx_v_tx2, __pyx_v_x2) - __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_max(__pyx_v_tx1, __pyx_v_x1)) + 1.0);
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":92
+ *             area = (x2 - x1 + 1) * (y2 - y1 + 1)
+ *             iw = (min(tx2, x2) - max(tx1, x1) + 1)
+ *             if iw > 0:             # <<<<<<<<<<<<<<
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ *                 if ih > 0:
+ */
+      __pyx_t_9 = ((__pyx_v_iw > 0.0) != 0);
+      if (__pyx_t_9) {
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":93
+ *             iw = (min(tx2, x2) - max(tx1, x1) + 1)
+ *             if iw > 0:
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)             # <<<<<<<<<<<<<<
+ *                 if ih > 0:
+ *                     ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)
+ */
+        __pyx_v_ih = ((__pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_min(__pyx_v_ty2, __pyx_v_y2) - __pyx_f_5mmdet_3ops_3nms_12soft_nms_cpu_max(__pyx_v_ty1, __pyx_v_y1)) + 1.0);
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":94
+ *             if iw > 0:
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ *                 if ih > 0:             # <<<<<<<<<<<<<<
+ *                     ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)
+ *                     ov = iw * ih / ua  # iou between max box and detection box
+ */
+        __pyx_t_9 = ((__pyx_v_ih > 0.0) != 0);
+        if (__pyx_t_9) {
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":95
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ *                 if ih > 0:
+ *                     ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)             # <<<<<<<<<<<<<<
+ *                     ov = iw * ih / ua  # iou between max box and detection box
+ * 
+ */
+          __pyx_v_ua = ((double)(((((__pyx_v_tx2 - __pyx_v_tx1) + 1.0) * ((__pyx_v_ty2 - __pyx_v_ty1) + 1.0)) + __pyx_v_area) - (__pyx_v_iw * __pyx_v_ih)));
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":96
+ *                 if ih > 0:
+ *                     ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)
+ *                     ov = iw * ih / ua  # iou between max box and detection box             # <<<<<<<<<<<<<<
+ * 
+ *                     if method == 1:  # linear
+ */
+          __pyx_t_8 = (__pyx_v_iw * __pyx_v_ih);
+          if (unlikely(__pyx_v_ua == 0)) {
+            PyErr_SetString(PyExc_ZeroDivisionError, "float division");
+            __PYX_ERR(0, 96, __pyx_L1_error)
+          }
+          __pyx_v_ov = (__pyx_t_8 / __pyx_v_ua);
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":98
+ *                     ov = iw * ih / ua  # iou between max box and detection box
+ * 
+ *                     if method == 1:  # linear             # <<<<<<<<<<<<<<
+ *                         if ov > iou_thr:
+ *                             weight = 1 - ov
+ */
+          switch (__pyx_v_method) {
+            case 1:
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":99
+ * 
+ *                     if method == 1:  # linear
+ *                         if ov > iou_thr:             # <<<<<<<<<<<<<<
+ *                             weight = 1 - ov
+ *                         else:
+ */
+            __pyx_t_9 = ((__pyx_v_ov > __pyx_v_iou_thr) != 0);
+            if (__pyx_t_9) {
+
+              /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":100
+ *                     if method == 1:  # linear
+ *                         if ov > iou_thr:
+ *                             weight = 1 - ov             # <<<<<<<<<<<<<<
+ *                         else:
+ *                             weight = 1
+ */
+              __pyx_v_weight = (1.0 - __pyx_v_ov);
+
+              /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":99
+ * 
+ *                     if method == 1:  # linear
+ *                         if ov > iou_thr:             # <<<<<<<<<<<<<<
+ *                             weight = 1 - ov
+ *                         else:
+ */
+              goto __pyx_L12;
+            }
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":102
+ *                             weight = 1 - ov
+ *                         else:
+ *                             weight = 1             # <<<<<<<<<<<<<<
+ *                     elif method == 2:  # gaussian
+ *                         weight = np.exp(-(ov * ov) / sigma)
+ */
+            /*else*/ {
+              __pyx_v_weight = 1.0;
+            }
+            __pyx_L12:;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":98
+ *                     ov = iw * ih / ua  # iou between max box and detection box
+ * 
+ *                     if method == 1:  # linear             # <<<<<<<<<<<<<<
+ *                         if ov > iou_thr:
+ *                             weight = 1 - ov
+ */
+            break;
+            case 2:
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":104
+ *                             weight = 1
+ *                     elif method == 2:  # gaussian
+ *                         weight = np.exp(-(ov * ov) / sigma)             # <<<<<<<<<<<<<<
+ *                     else:  # original NMS
+ *                         if ov > iou_thr:
+ */
+            __Pyx_GetModuleGlobalName(__pyx_t_5, __pyx_n_s_np); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 104, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_5, __pyx_n_s_exp); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 104, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __pyx_t_8 = (-(__pyx_v_ov * __pyx_v_ov));
+            if (unlikely(__pyx_v_sigma == 0)) {
+              PyErr_SetString(PyExc_ZeroDivisionError, "float division");
+              __PYX_ERR(0, 104, __pyx_L1_error)
+            }
+            __pyx_t_5 = PyFloat_FromDouble((__pyx_t_8 / __pyx_v_sigma)); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 104, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __pyx_t_10 = NULL;
+            if (CYTHON_UNPACK_METHODS && unlikely(PyMethod_Check(__pyx_t_1))) {
+              __pyx_t_10 = PyMethod_GET_SELF(__pyx_t_1);
+              if (likely(__pyx_t_10)) {
+                PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_1);
+                __Pyx_INCREF(__pyx_t_10);
+                __Pyx_INCREF(function);
+                __Pyx_DECREF_SET(__pyx_t_1, function);
+              }
+            }
+            __pyx_t_3 = (__pyx_t_10) ? __Pyx_PyObject_Call2Args(__pyx_t_1, __pyx_t_10, __pyx_t_5) : __Pyx_PyObject_CallOneArg(__pyx_t_1, __pyx_t_5);
+            __Pyx_XDECREF(__pyx_t_10); __pyx_t_10 = 0;
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 104, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+            __pyx_t_8 = __pyx_PyFloat_AsFloat(__pyx_t_3); if (unlikely((__pyx_t_8 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 104, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __pyx_v_weight = __pyx_t_8;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":103
+ *                         else:
+ *                             weight = 1
+ *                     elif method == 2:  # gaussian             # <<<<<<<<<<<<<<
+ *                         weight = np.exp(-(ov * ov) / sigma)
+ *                     else:  # original NMS
+ */
+            break;
+            default:
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":106
+ *                         weight = np.exp(-(ov * ov) / sigma)
+ *                     else:  # original NMS
+ *                         if ov > iou_thr:             # <<<<<<<<<<<<<<
+ *                             weight = 0
+ *                         else:
+ */
+            __pyx_t_9 = ((__pyx_v_ov > __pyx_v_iou_thr) != 0);
+            if (__pyx_t_9) {
+
+              /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":107
+ *                     else:  # original NMS
+ *                         if ov > iou_thr:
+ *                             weight = 0             # <<<<<<<<<<<<<<
+ *                         else:
+ *                             weight = 1
+ */
+              __pyx_v_weight = 0.0;
+
+              /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":106
+ *                         weight = np.exp(-(ov * ov) / sigma)
+ *                     else:  # original NMS
+ *                         if ov > iou_thr:             # <<<<<<<<<<<<<<
+ *                             weight = 0
+ *                         else:
+ */
+              goto __pyx_L13;
+            }
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":109
+ *                             weight = 0
+ *                         else:
+ *                             weight = 1             # <<<<<<<<<<<<<<
+ * 
+ *                     boxes[pos, 4] = weight * boxes[pos, 4]
+ */
+            /*else*/ {
+              __pyx_v_weight = 1.0;
+            }
+            __pyx_L13:;
+            break;
+          }
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":111
+ *                             weight = 1
+ * 
+ *                     boxes[pos, 4] = weight * boxes[pos, 4]             # <<<<<<<<<<<<<<
+ * 
+ *                     # if box score falls below threshold, discard the box by
+ */
+          __pyx_t_3 = PyFloat_FromDouble(__pyx_v_weight); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_3);
+          __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_1);
+          __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_5);
+          __Pyx_GIVEREF(__pyx_t_1);
+          PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_1);
+          __Pyx_INCREF(__pyx_int_4);
+          __Pyx_GIVEREF(__pyx_int_4);
+          PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_4);
+          __pyx_t_1 = 0;
+          __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_1);
+          __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+          __pyx_t_5 = PyNumber_Multiply(__pyx_t_3, __pyx_t_1); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_5);
+          __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+          __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+          __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_1);
+          __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_3);
+          __Pyx_GIVEREF(__pyx_t_1);
+          PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+          __Pyx_INCREF(__pyx_int_4);
+          __Pyx_GIVEREF(__pyx_int_4);
+          PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_4);
+          __pyx_t_1 = 0;
+          if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_5) < 0)) __PYX_ERR(0, 111, __pyx_L1_error)
+          __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+          __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":115
+ *                     # if box score falls below threshold, discard the box by
+ *                     # swapping with last box update N
+ *                     if boxes[pos, 4] < min_score:             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 0] = boxes[N-1, 0]
+ *                         boxes[pos, 1] = boxes[N-1, 1]
+ */
+          __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_5);
+          __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_3);
+          __Pyx_GIVEREF(__pyx_t_5);
+          PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_5);
+          __Pyx_INCREF(__pyx_int_4);
+          __Pyx_GIVEREF(__pyx_int_4);
+          PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_4);
+          __pyx_t_5 = 0;
+          __pyx_t_5 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_5);
+          __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+          __pyx_t_3 = PyFloat_FromDouble(__pyx_v_min_score); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_GOTREF(__pyx_t_3);
+          __pyx_t_1 = PyObject_RichCompare(__pyx_t_5, __pyx_t_3, Py_LT); __Pyx_XGOTREF(__pyx_t_1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+          __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+          __pyx_t_9 = __Pyx_PyObject_IsTrue(__pyx_t_1); if (unlikely(__pyx_t_9 < 0)) __PYX_ERR(0, 115, __pyx_L1_error)
+          __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+          if (__pyx_t_9) {
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":116
+ *                     # swapping with last box update N
+ *                     if boxes[pos, 4] < min_score:
+ *                         boxes[pos, 0] = boxes[N-1, 0]             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 1] = boxes[N-1, 1]
+ *                         boxes[pos, 2] = boxes[N-1, 2]
+ */
+            __pyx_t_1 = __Pyx_PyInt_From_long((__pyx_v_N - 1)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_GIVEREF(__pyx_t_1);
+            PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+            __Pyx_INCREF(__pyx_int_0);
+            __Pyx_GIVEREF(__pyx_int_0);
+            PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_0);
+            __pyx_t_1 = 0;
+            __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __Pyx_GIVEREF(__pyx_t_3);
+            PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+            __Pyx_INCREF(__pyx_int_0);
+            __Pyx_GIVEREF(__pyx_int_0);
+            PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_0);
+            __pyx_t_3 = 0;
+            if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_5, __pyx_t_1) < 0)) __PYX_ERR(0, 116, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":117
+ *                     if boxes[pos, 4] < min_score:
+ *                         boxes[pos, 0] = boxes[N-1, 0]
+ *                         boxes[pos, 1] = boxes[N-1, 1]             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 2] = boxes[N-1, 2]
+ *                         boxes[pos, 3] = boxes[N-1, 3]
+ */
+            __pyx_t_1 = __Pyx_PyInt_From_long((__pyx_v_N - 1)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __Pyx_GIVEREF(__pyx_t_1);
+            PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_1);
+            __Pyx_INCREF(__pyx_int_1);
+            __Pyx_GIVEREF(__pyx_int_1);
+            PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_1);
+            __pyx_t_1 = 0;
+            __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_GIVEREF(__pyx_t_5);
+            PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_5);
+            __Pyx_INCREF(__pyx_int_1);
+            __Pyx_GIVEREF(__pyx_int_1);
+            PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_1);
+            __pyx_t_5 = 0;
+            if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_1) < 0)) __PYX_ERR(0, 117, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":118
+ *                         boxes[pos, 0] = boxes[N-1, 0]
+ *                         boxes[pos, 1] = boxes[N-1, 1]
+ *                         boxes[pos, 2] = boxes[N-1, 2]             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 3] = boxes[N-1, 3]
+ *                         boxes[pos, 4] = boxes[N-1, 4]
+ */
+            __pyx_t_1 = __Pyx_PyInt_From_long((__pyx_v_N - 1)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_GIVEREF(__pyx_t_1);
+            PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+            __Pyx_INCREF(__pyx_int_2);
+            __Pyx_GIVEREF(__pyx_int_2);
+            PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_2);
+            __pyx_t_1 = 0;
+            __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __Pyx_GIVEREF(__pyx_t_3);
+            PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+            __Pyx_INCREF(__pyx_int_2);
+            __Pyx_GIVEREF(__pyx_int_2);
+            PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_2);
+            __pyx_t_3 = 0;
+            if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_5, __pyx_t_1) < 0)) __PYX_ERR(0, 118, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":119
+ *                         boxes[pos, 1] = boxes[N-1, 1]
+ *                         boxes[pos, 2] = boxes[N-1, 2]
+ *                         boxes[pos, 3] = boxes[N-1, 3]             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 4] = boxes[N-1, 4]
+ *                         inds[pos] = inds[N - 1]
+ */
+            __pyx_t_1 = __Pyx_PyInt_From_long((__pyx_v_N - 1)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __Pyx_GIVEREF(__pyx_t_1);
+            PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_1);
+            __Pyx_INCREF(__pyx_int_3);
+            __Pyx_GIVEREF(__pyx_int_3);
+            PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_3);
+            __pyx_t_1 = 0;
+            __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_5); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_GIVEREF(__pyx_t_5);
+            PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_5);
+            __Pyx_INCREF(__pyx_int_3);
+            __Pyx_GIVEREF(__pyx_int_3);
+            PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_3);
+            __pyx_t_5 = 0;
+            if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_3, __pyx_t_1) < 0)) __PYX_ERR(0, 119, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":120
+ *                         boxes[pos, 2] = boxes[N-1, 2]
+ *                         boxes[pos, 3] = boxes[N-1, 3]
+ *                         boxes[pos, 4] = boxes[N-1, 4]             # <<<<<<<<<<<<<<
+ *                         inds[pos] = inds[N - 1]
+ *                         N = N - 1
+ */
+            __pyx_t_1 = __Pyx_PyInt_From_long((__pyx_v_N - 1)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __Pyx_GIVEREF(__pyx_t_1);
+            PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
+            __Pyx_INCREF(__pyx_int_4);
+            __Pyx_GIVEREF(__pyx_int_4);
+            PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_4);
+            __pyx_t_1 = 0;
+            __pyx_t_1 = __Pyx_PyObject_GetItem(__pyx_v_boxes, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+            __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_pos); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_3);
+            __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_5);
+            __Pyx_GIVEREF(__pyx_t_3);
+            PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_3);
+            __Pyx_INCREF(__pyx_int_4);
+            __Pyx_GIVEREF(__pyx_int_4);
+            PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_4);
+            __pyx_t_3 = 0;
+            if (unlikely(PyObject_SetItem(__pyx_v_boxes, __pyx_t_5, __pyx_t_1) < 0)) __PYX_ERR(0, 120, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":121
+ *                         boxes[pos, 3] = boxes[N-1, 3]
+ *                         boxes[pos, 4] = boxes[N-1, 4]
+ *                         inds[pos] = inds[N - 1]             # <<<<<<<<<<<<<<
+ *                         N = N - 1
+ *                         pos = pos - 1
+ */
+            __pyx_t_11 = (__pyx_v_N - 1);
+            __pyx_t_1 = __Pyx_GetItemInt(__pyx_v_inds, __pyx_t_11, long, 1, __Pyx_PyInt_From_long, 0, 1, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 121, __pyx_L1_error)
+            __Pyx_GOTREF(__pyx_t_1);
+            if (unlikely(__Pyx_SetItemInt(__pyx_v_inds, __pyx_v_pos, __pyx_t_1, int, 1, __Pyx_PyInt_From_int, 0, 1, 0) < 0)) __PYX_ERR(0, 121, __pyx_L1_error)
+            __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":122
+ *                         boxes[pos, 4] = boxes[N-1, 4]
+ *                         inds[pos] = inds[N - 1]
+ *                         N = N - 1             # <<<<<<<<<<<<<<
+ *                         pos = pos - 1
+ * 
+ */
+            __pyx_v_N = (__pyx_v_N - 1);
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":123
+ *                         inds[pos] = inds[N - 1]
+ *                         N = N - 1
+ *                         pos = pos - 1             # <<<<<<<<<<<<<<
+ * 
+ *             pos = pos + 1
+ */
+            __pyx_v_pos = (__pyx_v_pos - 1);
+
+            /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":115
+ *                     # if box score falls below threshold, discard the box by
+ *                     # swapping with last box update N
+ *                     if boxes[pos, 4] < min_score:             # <<<<<<<<<<<<<<
+ *                         boxes[pos, 0] = boxes[N-1, 0]
+ *                         boxes[pos, 1] = boxes[N-1, 1]
+ */
+          }
+
+          /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":94
+ *             if iw > 0:
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ *                 if ih > 0:             # <<<<<<<<<<<<<<
+ *                     ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)
+ *                     ov = iw * ih / ua  # iou between max box and detection box
+ */
+        }
+
+        /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":92
+ *             area = (x2 - x1 + 1) * (y2 - y1 + 1)
+ *             iw = (min(tx2, x2) - max(tx1, x1) + 1)
+ *             if iw > 0:             # <<<<<<<<<<<<<<
+ *                 ih = (min(ty2, y2) - max(ty1, y1) + 1)
+ *                 if ih > 0:
+ */
+      }
+
+      /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":125
+ *                         pos = pos - 1
+ * 
+ *             pos = pos + 1             # <<<<<<<<<<<<<<
+ * 
+ *     return boxes[:N], inds[:N]
+ */
+      __pyx_v_pos = (__pyx_v_pos + 1);
+    }
+
+    /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":39
+ *     inds = np.arange(N)
+ * 
+ *     for i in range(N):             # <<<<<<<<<<<<<<
+ *         maxscore = boxes[i, 4]
+ *         maxpos = i
+ */
+  }
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":127
+ *             pos = pos + 1
+ * 
+ *     return boxes[:N], inds[:N]             # <<<<<<<<<<<<<<
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_2 = __Pyx_PyObject_GetSlice(__pyx_v_boxes, 0, __pyx_v_N, NULL, NULL, NULL, 0, 1, 1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 127, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __pyx_t_1 = __Pyx_PyObject_GetSlice(__pyx_v_inds, 0, __pyx_v_N, NULL, NULL, NULL, 0, 1, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 127, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_t_5 = PyTuple_New(2); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 127, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_5);
+  __Pyx_GIVEREF(__pyx_t_2);
+  PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_2);
+  __Pyx_GIVEREF(__pyx_t_1);
+  PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_1);
+  __pyx_t_2 = 0;
+  __pyx_t_1 = 0;
+  __pyx_r = __pyx_t_5;
+  __pyx_t_5 = 0;
+  goto __pyx_L0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":22
+ * 
+ * 
+ * def soft_nms_cpu(             # <<<<<<<<<<<<<<
+ *     np.ndarray[float, ndim=2] boxes_in,
+ *     float iou_thr,
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_XDECREF(__pyx_t_2);
+  __Pyx_XDECREF(__pyx_t_3);
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_10);
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_boxes_in.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mmdet.ops.nms.soft_nms_cpu.soft_nms_cpu", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_boxes_in.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XDECREF(__pyx_v_boxes);
+  __Pyx_XDECREF(__pyx_v_inds);
+  __Pyx_XDECREF(__pyx_v_i);
+  __Pyx_XDECREF(__pyx_v_ti);
+  __Pyx_XDECREF(__pyx_v_s);
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":258
+ *         # experimental exception made for __getbuffer__ and __releasebuffer__
+ *         # -- the details of this may change.
+ *         def __getbuffer__(ndarray self, Py_buffer* info, int flags):             # <<<<<<<<<<<<<<
+ *             # This implementation of getbuffer is geared towards Cython
+ *             # requirements, and does not yet fulfill the PEP.
+ */
+
+/* Python wrapper */
+static CYTHON_UNUSED int __pyx_pw_5numpy_7ndarray_1__getbuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/
+static CYTHON_UNUSED int __pyx_pw_5numpy_7ndarray_1__getbuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0);
+  __pyx_r = __pyx_pf_5numpy_7ndarray___getbuffer__(((PyArrayObject *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags));
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static int __pyx_pf_5numpy_7ndarray___getbuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) {
+  int __pyx_v_i;
+  int __pyx_v_ndim;
+  int __pyx_v_endian_detector;
+  int __pyx_v_little_endian;
+  int __pyx_v_t;
+  char *__pyx_v_f;
+  PyArray_Descr *__pyx_v_descr = 0;
+  int __pyx_v_offset;
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  int __pyx_t_2;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  int __pyx_t_5;
+  int __pyx_t_6;
+  PyArray_Descr *__pyx_t_7;
+  PyObject *__pyx_t_8 = NULL;
+  char *__pyx_t_9;
+  if (__pyx_v_info == NULL) {
+    PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete");
+    return -1;
+  }
+  __Pyx_RefNannySetupContext("__getbuffer__", 0);
+  __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None);
+  __Pyx_GIVEREF(__pyx_v_info->obj);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":265
+ * 
+ *             cdef int i, ndim
+ *             cdef int endian_detector = 1             # <<<<<<<<<<<<<<
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ * 
+ */
+  __pyx_v_endian_detector = 1;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":266
+ *             cdef int i, ndim
+ *             cdef int endian_detector = 1
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)             # <<<<<<<<<<<<<<
+ * 
+ *             ndim = PyArray_NDIM(self)
+ */
+  __pyx_v_little_endian = ((((char *)(&__pyx_v_endian_detector))[0]) != 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":268
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ * 
+ *             ndim = PyArray_NDIM(self)             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ */
+  __pyx_v_ndim = PyArray_NDIM(__pyx_v_self);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":270
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  __pyx_t_2 = (((__pyx_v_flags & PyBUF_C_CONTIGUOUS) == PyBUF_C_CONTIGUOUS) != 0);
+  if (__pyx_t_2) {
+  } else {
+    __pyx_t_1 = __pyx_t_2;
+    goto __pyx_L4_bool_binop_done;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":271
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):             # <<<<<<<<<<<<<<
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ */
+  __pyx_t_2 = ((!(PyArray_CHKFLAGS(__pyx_v_self, NPY_ARRAY_C_CONTIGUOUS) != 0)) != 0);
+  __pyx_t_1 = __pyx_t_2;
+  __pyx_L4_bool_binop_done:;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":270
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  if (unlikely(__pyx_t_1)) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":272
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ */
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple_, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 272, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __PYX_ERR(1, 272, __pyx_L1_error)
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":270
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":274
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  __pyx_t_2 = (((__pyx_v_flags & PyBUF_F_CONTIGUOUS) == PyBUF_F_CONTIGUOUS) != 0);
+  if (__pyx_t_2) {
+  } else {
+    __pyx_t_1 = __pyx_t_2;
+    goto __pyx_L7_bool_binop_done;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":275
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):             # <<<<<<<<<<<<<<
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ * 
+ */
+  __pyx_t_2 = ((!(PyArray_CHKFLAGS(__pyx_v_self, NPY_ARRAY_F_CONTIGUOUS) != 0)) != 0);
+  __pyx_t_1 = __pyx_t_2;
+  __pyx_L7_bool_binop_done:;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":274
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  if (unlikely(__pyx_t_1)) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":276
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             info.buf = PyArray_DATA(self)
+ */
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__2, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 276, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __PYX_ERR(1, 276, __pyx_L1_error)
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":274
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":278
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ * 
+ *             info.buf = PyArray_DATA(self)             # <<<<<<<<<<<<<<
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  __pyx_v_info->buf = PyArray_DATA(__pyx_v_self);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":279
+ * 
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim             # <<<<<<<<<<<<<<
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 # Allocate new buffer for strides and shape info.
+ */
+  __pyx_v_info->ndim = __pyx_v_ndim;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":280
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ */
+  __pyx_t_1 = (((sizeof(npy_intp)) != (sizeof(Py_ssize_t))) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":283
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)             # <<<<<<<<<<<<<<
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):
+ */
+    __pyx_v_info->strides = ((Py_ssize_t *)PyObject_Malloc((((sizeof(Py_ssize_t)) * 2) * ((size_t)__pyx_v_ndim))));
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":284
+ *                 # This is allocated as one block, strides first.
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)
+ *                 info.shape = info.strides + ndim             # <<<<<<<<<<<<<<
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ */
+    __pyx_v_info->shape = (__pyx_v_info->strides + __pyx_v_ndim);
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":285
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):             # <<<<<<<<<<<<<<
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ */
+    __pyx_t_4 = __pyx_v_ndim;
+    __pyx_t_5 = __pyx_t_4;
+    for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
+      __pyx_v_i = __pyx_t_6;
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":286
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]             # <<<<<<<<<<<<<<
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ *             else:
+ */
+      (__pyx_v_info->strides[__pyx_v_i]) = (PyArray_STRIDES(__pyx_v_self)[__pyx_v_i]);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":287
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ *                     info.shape[i] = PyArray_DIMS(self)[i]             # <<<<<<<<<<<<<<
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ */
+      (__pyx_v_info->shape[__pyx_v_i]) = (PyArray_DIMS(__pyx_v_self)[__pyx_v_i]);
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":280
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ */
+    goto __pyx_L9;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":289
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)             # <<<<<<<<<<<<<<
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL
+ */
+  /*else*/ {
+    __pyx_v_info->strides = ((Py_ssize_t *)PyArray_STRIDES(__pyx_v_self));
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":290
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)             # <<<<<<<<<<<<<<
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ */
+    __pyx_v_info->shape = ((Py_ssize_t *)PyArray_DIMS(__pyx_v_self));
+  }
+  __pyx_L9:;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":291
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL             # <<<<<<<<<<<<<<
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ *             info.readonly = not PyArray_ISWRITEABLE(self)
+ */
+  __pyx_v_info->suboffsets = NULL;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":292
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)             # <<<<<<<<<<<<<<
+ *             info.readonly = not PyArray_ISWRITEABLE(self)
+ * 
+ */
+  __pyx_v_info->itemsize = PyArray_ITEMSIZE(__pyx_v_self);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":293
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ *             info.readonly = not PyArray_ISWRITEABLE(self)             # <<<<<<<<<<<<<<
+ * 
+ *             cdef int t
+ */
+  __pyx_v_info->readonly = (!(PyArray_ISWRITEABLE(__pyx_v_self) != 0));
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":296
+ * 
+ *             cdef int t
+ *             cdef char* f = NULL             # <<<<<<<<<<<<<<
+ *             cdef dtype descr = <dtype>PyArray_DESCR(self)
+ *             cdef int offset
+ */
+  __pyx_v_f = NULL;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":297
+ *             cdef int t
+ *             cdef char* f = NULL
+ *             cdef dtype descr = <dtype>PyArray_DESCR(self)             # <<<<<<<<<<<<<<
+ *             cdef int offset
+ * 
+ */
+  __pyx_t_7 = PyArray_DESCR(__pyx_v_self);
+  __pyx_t_3 = ((PyObject *)__pyx_t_7);
+  __Pyx_INCREF(__pyx_t_3);
+  __pyx_v_descr = ((PyArray_Descr *)__pyx_t_3);
+  __pyx_t_3 = 0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":300
+ *             cdef int offset
+ * 
+ *             info.obj = self             # <<<<<<<<<<<<<<
+ * 
+ *             if not PyDataType_HASFIELDS(descr):
+ */
+  __Pyx_INCREF(((PyObject *)__pyx_v_self));
+  __Pyx_GIVEREF(((PyObject *)__pyx_v_self));
+  __Pyx_GOTREF(__pyx_v_info->obj);
+  __Pyx_DECREF(__pyx_v_info->obj);
+  __pyx_v_info->obj = ((PyObject *)__pyx_v_self);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":302
+ *             info.obj = self
+ * 
+ *             if not PyDataType_HASFIELDS(descr):             # <<<<<<<<<<<<<<
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ */
+  __pyx_t_1 = ((!(PyDataType_HASFIELDS(__pyx_v_descr) != 0)) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":303
+ * 
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num             # <<<<<<<<<<<<<<
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ */
+    __pyx_t_4 = __pyx_v_descr->type_num;
+    __pyx_v_t = __pyx_t_4;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":304
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    __pyx_t_2 = ((__pyx_v_descr->byteorder == '>') != 0);
+    if (!__pyx_t_2) {
+      goto __pyx_L15_next_or;
+    } else {
+    }
+    __pyx_t_2 = (__pyx_v_little_endian != 0);
+    if (!__pyx_t_2) {
+    } else {
+      __pyx_t_1 = __pyx_t_2;
+      goto __pyx_L14_bool_binop_done;
+    }
+    __pyx_L15_next_or:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":305
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):             # <<<<<<<<<<<<<<
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"
+ */
+    __pyx_t_2 = ((__pyx_v_descr->byteorder == '<') != 0);
+    if (__pyx_t_2) {
+    } else {
+      __pyx_t_1 = __pyx_t_2;
+      goto __pyx_L14_bool_binop_done;
+    }
+    __pyx_t_2 = ((!(__pyx_v_little_endian != 0)) != 0);
+    __pyx_t_1 = __pyx_t_2;
+    __pyx_L14_bool_binop_done:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":304
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    if (unlikely(__pyx_t_1)) {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":306
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__3, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 306, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 306, __pyx_L1_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":304
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":307
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"
+ */
+    switch (__pyx_v_t) {
+      case NPY_BYTE:
+      __pyx_v_f = ((char *)"b");
+      break;
+      case NPY_UBYTE:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":308
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"
+ */
+      __pyx_v_f = ((char *)"B");
+      break;
+      case NPY_SHORT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":309
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"
+ */
+      __pyx_v_f = ((char *)"h");
+      break;
+      case NPY_USHORT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":310
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"
+ */
+      __pyx_v_f = ((char *)"H");
+      break;
+      case NPY_INT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":311
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"
+ */
+      __pyx_v_f = ((char *)"i");
+      break;
+      case NPY_UINT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":312
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"
+ */
+      __pyx_v_f = ((char *)"I");
+      break;
+      case NPY_LONG:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":313
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ */
+      __pyx_v_f = ((char *)"l");
+      break;
+      case NPY_ULONG:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":314
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ */
+      __pyx_v_f = ((char *)"L");
+      break;
+      case NPY_LONGLONG:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":315
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"
+ */
+      __pyx_v_f = ((char *)"q");
+      break;
+      case NPY_ULONGLONG:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":316
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ */
+      __pyx_v_f = ((char *)"Q");
+      break;
+      case NPY_FLOAT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":317
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ */
+      __pyx_v_f = ((char *)"f");
+      break;
+      case NPY_DOUBLE:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":318
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ */
+      __pyx_v_f = ((char *)"d");
+      break;
+      case NPY_LONGDOUBLE:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":319
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ */
+      __pyx_v_f = ((char *)"g");
+      break;
+      case NPY_CFLOAT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":320
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ */
+      __pyx_v_f = ((char *)"Zf");
+      break;
+      case NPY_CDOUBLE:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":321
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ *                 elif t == NPY_OBJECT:      f = "O"
+ */
+      __pyx_v_f = ((char *)"Zd");
+      break;
+      case NPY_CLONGDOUBLE:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":322
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_OBJECT:      f = "O"
+ *                 else:
+ */
+      __pyx_v_f = ((char *)"Zg");
+      break;
+      case NPY_OBJECT:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":323
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ *                 elif t == NPY_OBJECT:      f = "O"             # <<<<<<<<<<<<<<
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ */
+      __pyx_v_f = ((char *)"O");
+      break;
+      default:
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":325
+ *                 elif t == NPY_OBJECT:      f = "O"
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)             # <<<<<<<<<<<<<<
+ *                 info.format = f
+ *                 return
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_t); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 325, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_8 = PyUnicode_Format(__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_t_3); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 325, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_8); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 325, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 325, __pyx_L1_error)
+      break;
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":326
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *                 info.format = f             # <<<<<<<<<<<<<<
+ *                 return
+ *             else:
+ */
+    __pyx_v_info->format = __pyx_v_f;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":327
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *                 info.format = f
+ *                 return             # <<<<<<<<<<<<<<
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ */
+    __pyx_r = 0;
+    goto __pyx_L0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":302
+ *             info.obj = self
+ * 
+ *             if not PyDataType_HASFIELDS(descr):             # <<<<<<<<<<<<<<
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":329
+ *                 return
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)             # <<<<<<<<<<<<<<
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0
+ */
+  /*else*/ {
+    __pyx_v_info->format = ((char *)PyObject_Malloc(0xFF));
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":330
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ *                 info.format[0] = c'^' # Native data types, manual alignment             # <<<<<<<<<<<<<<
+ *                 offset = 0
+ *                 f = _util_dtypestring(descr, info.format + 1,
+ */
+    (__pyx_v_info->format[0]) = '^';
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":331
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0             # <<<<<<<<<<<<<<
+ *                 f = _util_dtypestring(descr, info.format + 1,
+ *                                       info.format + _buffer_format_string_len,
+ */
+    __pyx_v_offset = 0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":332
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0
+ *                 f = _util_dtypestring(descr, info.format + 1,             # <<<<<<<<<<<<<<
+ *                                       info.format + _buffer_format_string_len,
+ *                                       &offset)
+ */
+    __pyx_t_9 = __pyx_f_5numpy__util_dtypestring(__pyx_v_descr, (__pyx_v_info->format + 1), (__pyx_v_info->format + 0xFF), (&__pyx_v_offset)); if (unlikely(__pyx_t_9 == ((char *)NULL))) __PYX_ERR(1, 332, __pyx_L1_error)
+    __pyx_v_f = __pyx_t_9;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":335
+ *                                       info.format + _buffer_format_string_len,
+ *                                       &offset)
+ *                 f[0] = c'\0' # Terminate format string             # <<<<<<<<<<<<<<
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ */
+    (__pyx_v_f[0]) = '\x00';
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":258
+ *         # experimental exception made for __getbuffer__ and __releasebuffer__
+ *         # -- the details of this may change.
+ *         def __getbuffer__(ndarray self, Py_buffer* info, int flags):             # <<<<<<<<<<<<<<
+ *             # This implementation of getbuffer is geared towards Cython
+ *             # requirements, and does not yet fulfill the PEP.
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_3);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.ndarray.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  if (__pyx_v_info->obj != NULL) {
+    __Pyx_GOTREF(__pyx_v_info->obj);
+    __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0;
+  }
+  goto __pyx_L2;
+  __pyx_L0:;
+  if (__pyx_v_info->obj == Py_None) {
+    __Pyx_GOTREF(__pyx_v_info->obj);
+    __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0;
+  }
+  __pyx_L2:;
+  __Pyx_XDECREF((PyObject *)__pyx_v_descr);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":337
+ *                 f[0] = c'\0' # Terminate format string
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):             # <<<<<<<<<<<<<<
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ */
+
+/* Python wrapper */
+static CYTHON_UNUSED void __pyx_pw_5numpy_7ndarray_3__releasebuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info); /*proto*/
+static CYTHON_UNUSED void __pyx_pw_5numpy_7ndarray_3__releasebuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__releasebuffer__ (wrapper)", 0);
+  __pyx_pf_5numpy_7ndarray_2__releasebuffer__(((PyArrayObject *)__pyx_v_self), ((Py_buffer *)__pyx_v_info));
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+static void __pyx_pf_5numpy_7ndarray_2__releasebuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info) {
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("__releasebuffer__", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":338
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  __pyx_t_1 = (PyArray_HASFIELDS(__pyx_v_self) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":339
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)             # <<<<<<<<<<<<<<
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 PyObject_Free(info.strides)
+ */
+    PyObject_Free(__pyx_v_info->format);
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":338
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":340
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.strides)
+ *                 # info.shape was stored after info.strides in the same block
+ */
+  __pyx_t_1 = (((sizeof(npy_intp)) != (sizeof(Py_ssize_t))) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":341
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 PyObject_Free(info.strides)             # <<<<<<<<<<<<<<
+ *                 # info.shape was stored after info.strides in the same block
+ * 
+ */
+    PyObject_Free(__pyx_v_info->strides);
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":340
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.strides)
+ *                 # info.shape was stored after info.strides in the same block
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":337
+ *                 f[0] = c'\0' # Terminate format string
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):             # <<<<<<<<<<<<<<
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ */
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":821
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew1(PyObject *__pyx_v_a) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew1", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":822
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ *     return PyArray_MultiIterNew(1, <void*>a)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(1, ((void *)__pyx_v_a)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 822, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":821
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew1", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":824
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew2(PyObject *__pyx_v_a, PyObject *__pyx_v_b) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew2", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":825
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(2, ((void *)__pyx_v_a), ((void *)__pyx_v_b)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 825, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":824
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew2", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":827
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew3(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew3", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":828
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(3, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 828, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":827
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew3", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":830
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew4(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew4", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":831
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(4, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 831, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":830
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew4", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":833
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew5(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d, PyObject *__pyx_v_e) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew5", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":834
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(5, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d), ((void *)__pyx_v_e)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 834, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":833
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew5", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":836
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyDataType_SHAPE(PyArray_Descr *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("PyDataType_SHAPE", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":837
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  __pyx_t_1 = (PyDataType_HASSUBARRAY(__pyx_v_d) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":838
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape             # <<<<<<<<<<<<<<
+ *     else:
+ *         return ()
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(((PyObject*)__pyx_v_d->subarray->shape));
+    __pyx_r = ((PyObject*)__pyx_v_d->subarray->shape);
+    goto __pyx_L0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":837
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":840
+ *         return <tuple>d.subarray.shape
+ *     else:
+ *         return ()             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:
+ */
+  /*else*/ {
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(__pyx_empty_tuple);
+    __pyx_r = __pyx_empty_tuple;
+    goto __pyx_L0;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":836
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":842
+ *         return ()
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:             # <<<<<<<<<<<<<<
+ *     # Recursive utility function used in __getbuffer__ to get format
+ *     # string. The new location in the format string is returned.
+ */
+
+static CYTHON_INLINE char *__pyx_f_5numpy__util_dtypestring(PyArray_Descr *__pyx_v_descr, char *__pyx_v_f, char *__pyx_v_end, int *__pyx_v_offset) {
+  PyArray_Descr *__pyx_v_child = 0;
+  int __pyx_v_endian_detector;
+  int __pyx_v_little_endian;
+  PyObject *__pyx_v_fields = 0;
+  PyObject *__pyx_v_childname = NULL;
+  PyObject *__pyx_v_new_offset = NULL;
+  PyObject *__pyx_v_t = NULL;
+  char *__pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  Py_ssize_t __pyx_t_2;
+  PyObject *__pyx_t_3 = NULL;
+  PyObject *__pyx_t_4 = NULL;
+  int __pyx_t_5;
+  int __pyx_t_6;
+  int __pyx_t_7;
+  long __pyx_t_8;
+  char *__pyx_t_9;
+  __Pyx_RefNannySetupContext("_util_dtypestring", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":847
+ * 
+ *     cdef dtype child
+ *     cdef int endian_detector = 1             # <<<<<<<<<<<<<<
+ *     cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ *     cdef tuple fields
+ */
+  __pyx_v_endian_detector = 1;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":848
+ *     cdef dtype child
+ *     cdef int endian_detector = 1
+ *     cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)             # <<<<<<<<<<<<<<
+ *     cdef tuple fields
+ * 
+ */
+  __pyx_v_little_endian = ((((char *)(&__pyx_v_endian_detector))[0]) != 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":851
+ *     cdef tuple fields
+ * 
+ *     for childname in descr.names:             # <<<<<<<<<<<<<<
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields
+ */
+  if (unlikely(__pyx_v_descr->names == Py_None)) {
+    PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable");
+    __PYX_ERR(1, 851, __pyx_L1_error)
+  }
+  __pyx_t_1 = __pyx_v_descr->names; __Pyx_INCREF(__pyx_t_1); __pyx_t_2 = 0;
+  for (;;) {
+    if (__pyx_t_2 >= PyTuple_GET_SIZE(__pyx_t_1)) break;
+    #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+    __pyx_t_3 = PyTuple_GET_ITEM(__pyx_t_1, __pyx_t_2); __Pyx_INCREF(__pyx_t_3); __pyx_t_2++; if (unlikely(0 < 0)) __PYX_ERR(1, 851, __pyx_L1_error)
+    #else
+    __pyx_t_3 = PySequence_ITEM(__pyx_t_1, __pyx_t_2); __pyx_t_2++; if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 851, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    #endif
+    __Pyx_XDECREF_SET(__pyx_v_childname, __pyx_t_3);
+    __pyx_t_3 = 0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":852
+ * 
+ *     for childname in descr.names:
+ *         fields = descr.fields[childname]             # <<<<<<<<<<<<<<
+ *         child, new_offset = fields
+ * 
+ */
+    if (unlikely(__pyx_v_descr->fields == Py_None)) {
+      PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable");
+      __PYX_ERR(1, 852, __pyx_L1_error)
+    }
+    __pyx_t_3 = __Pyx_PyDict_GetItem(__pyx_v_descr->fields, __pyx_v_childname); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 852, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    if (!(likely(PyTuple_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "tuple", Py_TYPE(__pyx_t_3)->tp_name), 0))) __PYX_ERR(1, 852, __pyx_L1_error)
+    __Pyx_XDECREF_SET(__pyx_v_fields, ((PyObject*)__pyx_t_3));
+    __pyx_t_3 = 0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":853
+ *     for childname in descr.names:
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields             # <<<<<<<<<<<<<<
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ */
+    if (likely(__pyx_v_fields != Py_None)) {
+      PyObject* sequence = __pyx_v_fields;
+      Py_ssize_t size = __Pyx_PySequence_SIZE(sequence);
+      if (unlikely(size != 2)) {
+        if (size > 2) __Pyx_RaiseTooManyValuesError(2);
+        else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size);
+        __PYX_ERR(1, 853, __pyx_L1_error)
+      }
+      #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+      __pyx_t_3 = PyTuple_GET_ITEM(sequence, 0); 
+      __pyx_t_4 = PyTuple_GET_ITEM(sequence, 1); 
+      __Pyx_INCREF(__pyx_t_3);
+      __Pyx_INCREF(__pyx_t_4);
+      #else
+      __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 853, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 853, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      #endif
+    } else {
+      __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 853, __pyx_L1_error)
+    }
+    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_ptype_5numpy_dtype))))) __PYX_ERR(1, 853, __pyx_L1_error)
+    __Pyx_XDECREF_SET(__pyx_v_child, ((PyArray_Descr *)__pyx_t_3));
+    __pyx_t_3 = 0;
+    __Pyx_XDECREF_SET(__pyx_v_new_offset, __pyx_t_4);
+    __pyx_t_4 = 0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":855
+ *         child, new_offset = fields
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:             # <<<<<<<<<<<<<<
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyInt_From_int((__pyx_v_offset[0])); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 855, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_4);
+    __pyx_t_3 = PyNumber_Subtract(__pyx_v_new_offset, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 855, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+    __pyx_t_5 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_5 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 855, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_6 = ((((__pyx_v_end - __pyx_v_f) - ((int)__pyx_t_5)) < 15) != 0);
+    if (unlikely(__pyx_t_6)) {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":856
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")             # <<<<<<<<<<<<<<
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_RuntimeError, __pyx_tuple__4, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 856, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 856, __pyx_L1_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":855
+ *         child, new_offset = fields
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:             # <<<<<<<<<<<<<<
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ */
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":858
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    __pyx_t_7 = ((__pyx_v_child->byteorder == '>') != 0);
+    if (!__pyx_t_7) {
+      goto __pyx_L8_next_or;
+    } else {
+    }
+    __pyx_t_7 = (__pyx_v_little_endian != 0);
+    if (!__pyx_t_7) {
+    } else {
+      __pyx_t_6 = __pyx_t_7;
+      goto __pyx_L7_bool_binop_done;
+    }
+    __pyx_L8_next_or:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":859
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ *             (child.byteorder == c'<' and not little_endian)):             # <<<<<<<<<<<<<<
+ *             raise ValueError(u"Non-native byte order not supported")
+ *             # One could encode it in the format string and have Cython
+ */
+    __pyx_t_7 = ((__pyx_v_child->byteorder == '<') != 0);
+    if (__pyx_t_7) {
+    } else {
+      __pyx_t_6 = __pyx_t_7;
+      goto __pyx_L7_bool_binop_done;
+    }
+    __pyx_t_7 = ((!(__pyx_v_little_endian != 0)) != 0);
+    __pyx_t_6 = __pyx_t_7;
+    __pyx_L7_bool_binop_done:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":858
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    if (unlikely(__pyx_t_6)) {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":860
+ *         if ((child.byteorder == c'>' and little_endian) or
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *             # One could encode it in the format string and have Cython
+ *             # complain instead, BUT: < and > in format strings also imply
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__3, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 860, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 860, __pyx_L1_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":858
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":870
+ * 
+ *         # Output padding bytes
+ *         while offset[0] < new_offset:             # <<<<<<<<<<<<<<
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1
+ */
+    while (1) {
+      __pyx_t_3 = __Pyx_PyInt_From_int((__pyx_v_offset[0])); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 870, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_t_3, __pyx_v_new_offset, Py_LT); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 870, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 870, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (!__pyx_t_6) break;
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":871
+ *         # Output padding bytes
+ *         while offset[0] < new_offset:
+ *             f[0] = 120 # "x"; pad byte             # <<<<<<<<<<<<<<
+ *             f += 1
+ *             offset[0] += 1
+ */
+      (__pyx_v_f[0]) = 0x78;
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":872
+ *         while offset[0] < new_offset:
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1             # <<<<<<<<<<<<<<
+ *             offset[0] += 1
+ * 
+ */
+      __pyx_v_f = (__pyx_v_f + 1);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":873
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1
+ *             offset[0] += 1             # <<<<<<<<<<<<<<
+ * 
+ *         offset[0] += child.itemsize
+ */
+      __pyx_t_8 = 0;
+      (__pyx_v_offset[__pyx_t_8]) = ((__pyx_v_offset[__pyx_t_8]) + 1);
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":875
+ *             offset[0] += 1
+ * 
+ *         offset[0] += child.itemsize             # <<<<<<<<<<<<<<
+ * 
+ *         if not PyDataType_HASFIELDS(child):
+ */
+    __pyx_t_8 = 0;
+    (__pyx_v_offset[__pyx_t_8]) = ((__pyx_v_offset[__pyx_t_8]) + __pyx_v_child->elsize);
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":877
+ *         offset[0] += child.itemsize
+ * 
+ *         if not PyDataType_HASFIELDS(child):             # <<<<<<<<<<<<<<
+ *             t = child.type_num
+ *             if end - f < 5:
+ */
+    __pyx_t_6 = ((!(PyDataType_HASFIELDS(__pyx_v_child) != 0)) != 0);
+    if (__pyx_t_6) {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":878
+ * 
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num             # <<<<<<<<<<<<<<
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_int(__pyx_v_child->type_num); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 878, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __Pyx_XDECREF_SET(__pyx_v_t, __pyx_t_4);
+      __pyx_t_4 = 0;
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":879
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num
+ *             if end - f < 5:             # <<<<<<<<<<<<<<
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ * 
+ */
+      __pyx_t_6 = (((__pyx_v_end - __pyx_v_f) < 5) != 0);
+      if (unlikely(__pyx_t_6)) {
+
+        /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":880
+ *             t = child.type_num
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")             # <<<<<<<<<<<<<<
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ */
+        __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin_RuntimeError, __pyx_tuple__5, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 880, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_4);
+        __Pyx_Raise(__pyx_t_4, 0, 0, 0);
+        __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+        __PYX_ERR(1, 880, __pyx_L1_error)
+
+        /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":879
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num
+ *             if end - f < 5:             # <<<<<<<<<<<<<<
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ * 
+ */
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":883
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_BYTE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 883, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 883, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 883, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 98;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":884
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_UBYTE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 884, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 884, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 884, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 66;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":885
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_SHORT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 885, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 885, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 885, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x68;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":886
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_USHORT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 886, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 886, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 886, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 72;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":887
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_INT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 887, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 887, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 887, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x69;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":888
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_UINT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 888, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 888, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 888, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 73;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":889
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONG); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 889, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 889, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 889, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x6C;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":890
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_ULONG); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 890, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 890, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 890, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 76;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":891
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONGLONG); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 891, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 891, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 891, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x71;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":892
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_ULONGLONG); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 892, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 892, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 892, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 81;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":893
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_FLOAT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 893, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 893, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 893, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x66;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":894
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_DOUBLE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 894, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 894, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 894, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x64;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":895
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONGDOUBLE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 895, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 895, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 895, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x67;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":896
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CFLOAT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 896, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 896, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 896, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x66;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":897
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CDOUBLE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 897, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 897, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 897, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x64;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":898
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg             # <<<<<<<<<<<<<<
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ *             else:
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CLONGDOUBLE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 898, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 898, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 898, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x67;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":899
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"             # <<<<<<<<<<<<<<
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_OBJECT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 899, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 899, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 899, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (likely(__pyx_t_6)) {
+        (__pyx_v_f[0]) = 79;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":901
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)             # <<<<<<<<<<<<<<
+ *             f += 1
+ *         else:
+ */
+      /*else*/ {
+        __pyx_t_3 = __Pyx_PyUnicode_FormatSafe(__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_v_t); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 901, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_3);
+        __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 901, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_4);
+        __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+        __Pyx_Raise(__pyx_t_4, 0, 0, 0);
+        __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+        __PYX_ERR(1, 901, __pyx_L1_error)
+      }
+      __pyx_L15:;
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":902
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *             f += 1             # <<<<<<<<<<<<<<
+ *         else:
+ *             # Cython ignores struct boundary information ("T{...}"),
+ */
+      __pyx_v_f = (__pyx_v_f + 1);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":877
+ *         offset[0] += child.itemsize
+ * 
+ *         if not PyDataType_HASFIELDS(child):             # <<<<<<<<<<<<<<
+ *             t = child.type_num
+ *             if end - f < 5:
+ */
+      goto __pyx_L13;
+    }
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":906
+ *             # Cython ignores struct boundary information ("T{...}"),
+ *             # so don't output it
+ *             f = _util_dtypestring(child, f, end, offset)             # <<<<<<<<<<<<<<
+ *     return f
+ * 
+ */
+    /*else*/ {
+      __pyx_t_9 = __pyx_f_5numpy__util_dtypestring(__pyx_v_child, __pyx_v_f, __pyx_v_end, __pyx_v_offset); if (unlikely(__pyx_t_9 == ((char *)NULL))) __PYX_ERR(1, 906, __pyx_L1_error)
+      __pyx_v_f = __pyx_t_9;
+    }
+    __pyx_L13:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":851
+ *     cdef tuple fields
+ * 
+ *     for childname in descr.names:             # <<<<<<<<<<<<<<
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields
+ */
+  }
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":907
+ *             # so don't output it
+ *             f = _util_dtypestring(child, f, end, offset)
+ *     return f             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = __pyx_v_f;
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":842
+ *         return ()
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:             # <<<<<<<<<<<<<<
+ *     # Recursive utility function used in __getbuffer__ to get format
+ *     # string. The new location in the format string is returned.
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_XDECREF(__pyx_t_3);
+  __Pyx_XDECREF(__pyx_t_4);
+  __Pyx_AddTraceback("numpy._util_dtypestring", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_XDECREF((PyObject *)__pyx_v_child);
+  __Pyx_XDECREF(__pyx_v_fields);
+  __Pyx_XDECREF(__pyx_v_childname);
+  __Pyx_XDECREF(__pyx_v_new_offset);
+  __Pyx_XDECREF(__pyx_v_t);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1022
+ *     int _import_umath() except -1
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)
+ */
+
+static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_arr, PyObject *__pyx_v_base) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("set_array_base", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1023
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *     Py_INCREF(base) # important to do this before stealing the reference below!             # <<<<<<<<<<<<<<
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ */
+  Py_INCREF(__pyx_v_base);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1024
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ */
+  (void)(PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_base));
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1022
+ *     int _import_umath() except -1
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)
+ */
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1026
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__pyx_v_arr) {
+  PyObject *__pyx_v_base;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("get_array_base", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1027
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)             # <<<<<<<<<<<<<<
+ *     if base is NULL:
+ *         return None
+ */
+  __pyx_v_base = PyArray_BASE(__pyx_v_arr);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1028
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     return <object>base
+ */
+  __pyx_t_1 = ((__pyx_v_base == NULL) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1029
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ *         return None             # <<<<<<<<<<<<<<
+ *     return <object>base
+ * 
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+    goto __pyx_L0;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1028
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     return <object>base
+ */
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1030
+ *     if base is NULL:
+ *         return None
+ *     return <object>base             # <<<<<<<<<<<<<<
+ * 
+ * # Versions of the import_* functions which are more suitable for
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __Pyx_INCREF(((PyObject *)__pyx_v_base));
+  __pyx_r = ((PyObject *)__pyx_v_base);
+  goto __pyx_L0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1026
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1034
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_array()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_array(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_array", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1035
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1036
+ * cdef inline int import_array() except -1:
+ *     try:
+ *         _import_array()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ */
+      __pyx_t_4 = _import_array(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 1036, __pyx_L3_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1035
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1037
+ *     try:
+ *         _import_array()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 1037, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1038
+ *         _import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__6, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1038, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1038, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1035
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1034
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_array()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1040
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_umath(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_umath", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1041
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1042
+ * cdef inline int import_umath() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 1042, __pyx_L3_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1041
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1043
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 1043, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1044
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__7, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1044, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1044, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1041
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1040
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1046
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_ufunc(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_ufunc", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1047
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1048
+ * cdef inline int import_ufunc() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 1048, __pyx_L3_error)
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1047
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1049
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 1049, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1050
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__7, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1050, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1050, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1047
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1046
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyMethodDef __pyx_methods[] = {
+  {0, 0, 0, 0}
+};
+
+#if PY_MAJOR_VERSION >= 3
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/
+static int __pyx_pymod_exec_soft_nms_cpu(PyObject* module); /*proto*/
+static PyModuleDef_Slot __pyx_moduledef_slots[] = {
+  {Py_mod_create, (void*)__pyx_pymod_create},
+  {Py_mod_exec, (void*)__pyx_pymod_exec_soft_nms_cpu},
+  {0, NULL}
+};
+#endif
+
+static struct PyModuleDef __pyx_moduledef = {
+    PyModuleDef_HEAD_INIT,
+    "soft_nms_cpu",
+    0, /* m_doc */
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    0, /* m_size */
+  #else
+    -1, /* m_size */
+  #endif
+    __pyx_methods /* m_methods */,
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    __pyx_moduledef_slots, /* m_slots */
+  #else
+    NULL, /* m_reload */
+  #endif
+    NULL, /* m_traverse */
+    NULL, /* m_clear */
+    NULL /* m_free */
+};
+#endif
+#ifndef CYTHON_SMALL_CODE
+#if defined(__clang__)
+    #define CYTHON_SMALL_CODE
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+    #define CYTHON_SMALL_CODE __attribute__((cold))
+#else
+    #define CYTHON_SMALL_CODE
+#endif
+#endif
+
+static __Pyx_StringTabEntry __pyx_string_tab[] = {
+  {&__pyx_kp_u_Format_string_allocated_too_shor, __pyx_k_Format_string_allocated_too_shor, sizeof(__pyx_k_Format_string_allocated_too_shor), 0, 1, 0, 0},
+  {&__pyx_kp_u_Format_string_allocated_too_shor_2, __pyx_k_Format_string_allocated_too_shor_2, sizeof(__pyx_k_Format_string_allocated_too_shor_2), 0, 1, 0, 0},
+  {&__pyx_n_s_ImportError, __pyx_k_ImportError, sizeof(__pyx_k_ImportError), 0, 0, 1, 1},
+  {&__pyx_n_s_N, __pyx_k_N, sizeof(__pyx_k_N), 0, 0, 1, 1},
+  {&__pyx_kp_u_Non_native_byte_order_not_suppor, __pyx_k_Non_native_byte_order_not_suppor, sizeof(__pyx_k_Non_native_byte_order_not_suppor), 0, 1, 0, 0},
+  {&__pyx_n_s_RuntimeError, __pyx_k_RuntimeError, sizeof(__pyx_k_RuntimeError), 0, 0, 1, 1},
+  {&__pyx_n_s_ValueError, __pyx_k_ValueError, sizeof(__pyx_k_ValueError), 0, 0, 1, 1},
+  {&__pyx_n_s_arange, __pyx_k_arange, sizeof(__pyx_k_arange), 0, 0, 1, 1},
+  {&__pyx_n_s_area, __pyx_k_area, sizeof(__pyx_k_area), 0, 0, 1, 1},
+  {&__pyx_n_s_box_area, __pyx_k_box_area, sizeof(__pyx_k_box_area), 0, 0, 1, 1},
+  {&__pyx_n_s_boxes, __pyx_k_boxes, sizeof(__pyx_k_boxes), 0, 0, 1, 1},
+  {&__pyx_n_s_boxes_in, __pyx_k_boxes_in, sizeof(__pyx_k_boxes_in), 0, 0, 1, 1},
+  {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1},
+  {&__pyx_n_s_copy, __pyx_k_copy, sizeof(__pyx_k_copy), 0, 0, 1, 1},
+  {&__pyx_n_s_exp, __pyx_k_exp, sizeof(__pyx_k_exp), 0, 0, 1, 1},
+  {&__pyx_n_s_i, __pyx_k_i, sizeof(__pyx_k_i), 0, 0, 1, 1},
+  {&__pyx_n_s_ih, __pyx_k_ih, sizeof(__pyx_k_ih), 0, 0, 1, 1},
+  {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1},
+  {&__pyx_n_s_inds, __pyx_k_inds, sizeof(__pyx_k_inds), 0, 0, 1, 1},
+  {&__pyx_n_s_iou_thr, __pyx_k_iou_thr, sizeof(__pyx_k_iou_thr), 0, 0, 1, 1},
+  {&__pyx_n_s_iw, __pyx_k_iw, sizeof(__pyx_k_iw), 0, 0, 1, 1},
+  {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1},
+  {&__pyx_n_s_maxpos, __pyx_k_maxpos, sizeof(__pyx_k_maxpos), 0, 0, 1, 1},
+  {&__pyx_n_s_maxscore, __pyx_k_maxscore, sizeof(__pyx_k_maxscore), 0, 0, 1, 1},
+  {&__pyx_n_s_method, __pyx_k_method, sizeof(__pyx_k_method), 0, 0, 1, 1},
+  {&__pyx_n_s_min_score, __pyx_k_min_score, sizeof(__pyx_k_min_score), 0, 0, 1, 1},
+  {&__pyx_n_s_mmdet_ops_nms_soft_nms_cpu, __pyx_k_mmdet_ops_nms_soft_nms_cpu, sizeof(__pyx_k_mmdet_ops_nms_soft_nms_cpu), 0, 0, 1, 1},
+  {&__pyx_kp_s_mmdet_ops_nms_src_soft_nms_cpu_p, __pyx_k_mmdet_ops_nms_src_soft_nms_cpu_p, sizeof(__pyx_k_mmdet_ops_nms_src_soft_nms_cpu_p), 0, 0, 1, 0},
+  {&__pyx_n_s_name, __pyx_k_name, sizeof(__pyx_k_name), 0, 0, 1, 1},
+  {&__pyx_kp_u_ndarray_is_not_C_contiguous, __pyx_k_ndarray_is_not_C_contiguous, sizeof(__pyx_k_ndarray_is_not_C_contiguous), 0, 1, 0, 0},
+  {&__pyx_kp_u_ndarray_is_not_Fortran_contiguou, __pyx_k_ndarray_is_not_Fortran_contiguou, sizeof(__pyx_k_ndarray_is_not_Fortran_contiguou), 0, 1, 0, 0},
+  {&__pyx_n_s_np, __pyx_k_np, sizeof(__pyx_k_np), 0, 0, 1, 1},
+  {&__pyx_n_s_numpy, __pyx_k_numpy, sizeof(__pyx_k_numpy), 0, 0, 1, 1},
+  {&__pyx_kp_u_numpy_core_multiarray_failed_to, __pyx_k_numpy_core_multiarray_failed_to, sizeof(__pyx_k_numpy_core_multiarray_failed_to), 0, 1, 0, 0},
+  {&__pyx_kp_u_numpy_core_umath_failed_to_impor, __pyx_k_numpy_core_umath_failed_to_impor, sizeof(__pyx_k_numpy_core_umath_failed_to_impor), 0, 1, 0, 0},
+  {&__pyx_n_s_ov, __pyx_k_ov, sizeof(__pyx_k_ov), 0, 0, 1, 1},
+  {&__pyx_n_s_pos, __pyx_k_pos, sizeof(__pyx_k_pos), 0, 0, 1, 1},
+  {&__pyx_n_s_range, __pyx_k_range, sizeof(__pyx_k_range), 0, 0, 1, 1},
+  {&__pyx_n_s_s, __pyx_k_s, sizeof(__pyx_k_s), 0, 0, 1, 1},
+  {&__pyx_n_s_shape, __pyx_k_shape, sizeof(__pyx_k_shape), 0, 0, 1, 1},
+  {&__pyx_n_s_sigma, __pyx_k_sigma, sizeof(__pyx_k_sigma), 0, 0, 1, 1},
+  {&__pyx_n_s_soft_nms_cpu, __pyx_k_soft_nms_cpu, sizeof(__pyx_k_soft_nms_cpu), 0, 0, 1, 1},
+  {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1},
+  {&__pyx_n_s_ti, __pyx_k_ti, sizeof(__pyx_k_ti), 0, 0, 1, 1},
+  {&__pyx_n_s_ts, __pyx_k_ts, sizeof(__pyx_k_ts), 0, 0, 1, 1},
+  {&__pyx_n_s_tx1, __pyx_k_tx1, sizeof(__pyx_k_tx1), 0, 0, 1, 1},
+  {&__pyx_n_s_tx2, __pyx_k_tx2, sizeof(__pyx_k_tx2), 0, 0, 1, 1},
+  {&__pyx_n_s_ty1, __pyx_k_ty1, sizeof(__pyx_k_ty1), 0, 0, 1, 1},
+  {&__pyx_n_s_ty2, __pyx_k_ty2, sizeof(__pyx_k_ty2), 0, 0, 1, 1},
+  {&__pyx_n_s_ua, __pyx_k_ua, sizeof(__pyx_k_ua), 0, 0, 1, 1},
+  {&__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_k_unknown_dtype_code_in_numpy_pxd, sizeof(__pyx_k_unknown_dtype_code_in_numpy_pxd), 0, 1, 0, 0},
+  {&__pyx_n_s_weight, __pyx_k_weight, sizeof(__pyx_k_weight), 0, 0, 1, 1},
+  {&__pyx_n_s_x1, __pyx_k_x1, sizeof(__pyx_k_x1), 0, 0, 1, 1},
+  {&__pyx_n_s_x2, __pyx_k_x2, sizeof(__pyx_k_x2), 0, 0, 1, 1},
+  {&__pyx_n_s_y1, __pyx_k_y1, sizeof(__pyx_k_y1), 0, 0, 1, 1},
+  {&__pyx_n_s_y2, __pyx_k_y2, sizeof(__pyx_k_y2), 0, 0, 1, 1},
+  {0, 0, 0, 0, 0, 0, 0}
+};
+static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) {
+  __pyx_builtin_range = __Pyx_GetBuiltinName(__pyx_n_s_range); if (!__pyx_builtin_range) __PYX_ERR(0, 39, __pyx_L1_error)
+  __pyx_builtin_ValueError = __Pyx_GetBuiltinName(__pyx_n_s_ValueError); if (!__pyx_builtin_ValueError) __PYX_ERR(1, 272, __pyx_L1_error)
+  __pyx_builtin_RuntimeError = __Pyx_GetBuiltinName(__pyx_n_s_RuntimeError); if (!__pyx_builtin_RuntimeError) __PYX_ERR(1, 856, __pyx_L1_error)
+  __pyx_builtin_ImportError = __Pyx_GetBuiltinName(__pyx_n_s_ImportError); if (!__pyx_builtin_ImportError) __PYX_ERR(1, 1038, __pyx_L1_error)
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":272
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ */
+  __pyx_tuple_ = PyTuple_Pack(1, __pyx_kp_u_ndarray_is_not_C_contiguous); if (unlikely(!__pyx_tuple_)) __PYX_ERR(1, 272, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple_);
+  __Pyx_GIVEREF(__pyx_tuple_);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":276
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_ARRAY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             info.buf = PyArray_DATA(self)
+ */
+  __pyx_tuple__2 = PyTuple_Pack(1, __pyx_kp_u_ndarray_is_not_Fortran_contiguou); if (unlikely(!__pyx_tuple__2)) __PYX_ERR(1, 276, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__2);
+  __Pyx_GIVEREF(__pyx_tuple__2);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":306
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ */
+  __pyx_tuple__3 = PyTuple_Pack(1, __pyx_kp_u_Non_native_byte_order_not_suppor); if (unlikely(!__pyx_tuple__3)) __PYX_ERR(1, 306, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__3);
+  __Pyx_GIVEREF(__pyx_tuple__3);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":856
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")             # <<<<<<<<<<<<<<
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ */
+  __pyx_tuple__4 = PyTuple_Pack(1, __pyx_kp_u_Format_string_allocated_too_shor); if (unlikely(!__pyx_tuple__4)) __PYX_ERR(1, 856, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__4);
+  __Pyx_GIVEREF(__pyx_tuple__4);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":880
+ *             t = child.type_num
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")             # <<<<<<<<<<<<<<
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ */
+  __pyx_tuple__5 = PyTuple_Pack(1, __pyx_kp_u_Format_string_allocated_too_shor_2); if (unlikely(!__pyx_tuple__5)) __PYX_ERR(1, 880, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__5);
+  __Pyx_GIVEREF(__pyx_tuple__5);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1038
+ *         _import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+  __pyx_tuple__6 = PyTuple_Pack(1, __pyx_kp_u_numpy_core_multiarray_failed_to); if (unlikely(!__pyx_tuple__6)) __PYX_ERR(1, 1038, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__6);
+  __Pyx_GIVEREF(__pyx_tuple__6);
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1044
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+  __pyx_tuple__7 = PyTuple_Pack(1, __pyx_kp_u_numpy_core_umath_failed_to_impor); if (unlikely(!__pyx_tuple__7)) __PYX_ERR(1, 1044, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__7);
+  __Pyx_GIVEREF(__pyx_tuple__7);
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":22
+ * 
+ * 
+ * def soft_nms_cpu(             # <<<<<<<<<<<<<<
+ *     np.ndarray[float, ndim=2] boxes_in,
+ *     float iou_thr,
+ */
+  __pyx_tuple__8 = PyTuple_Pack(30, __pyx_n_s_boxes_in, __pyx_n_s_iou_thr, __pyx_n_s_method, __pyx_n_s_sigma, __pyx_n_s_min_score, __pyx_n_s_boxes, __pyx_n_s_N, __pyx_n_s_iw, __pyx_n_s_ih, __pyx_n_s_box_area, __pyx_n_s_ua, __pyx_n_s_pos, __pyx_n_s_maxscore, __pyx_n_s_maxpos, __pyx_n_s_x1, __pyx_n_s_x2, __pyx_n_s_y1, __pyx_n_s_y2, __pyx_n_s_tx1, __pyx_n_s_tx2, __pyx_n_s_ty1, __pyx_n_s_ty2, __pyx_n_s_ts, __pyx_n_s_area, __pyx_n_s_weight, __pyx_n_s_ov, __pyx_n_s_inds, __pyx_n_s_i, __pyx_n_s_ti, __pyx_n_s_s); if (unlikely(!__pyx_tuple__8)) __PYX_ERR(0, 22, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__8);
+  __Pyx_GIVEREF(__pyx_tuple__8);
+  __pyx_codeobj__9 = (PyObject*)__Pyx_PyCode_New(5, 0, 30, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__8, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_mmdet_ops_nms_src_soft_nms_cpu_p, __pyx_n_s_soft_nms_cpu, 22, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__9)) __PYX_ERR(0, 22, __pyx_L1_error)
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitGlobals(void) {
+  if (__Pyx_InitStrings(__pyx_string_tab) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  __pyx_int_0 = PyInt_FromLong(0); if (unlikely(!__pyx_int_0)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_int_1 = PyInt_FromLong(1); if (unlikely(!__pyx_int_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_int_2 = PyInt_FromLong(2); if (unlikely(!__pyx_int_2)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_int_3 = PyInt_FromLong(3); if (unlikely(!__pyx_int_3)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_int_4 = PyInt_FromLong(4); if (unlikely(!__pyx_int_4)) __PYX_ERR(0, 1, __pyx_L1_error)
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_modinit_global_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_import_code(void); /*proto*/
+
+static int __Pyx_modinit_global_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0);
+  /*--- Global init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_variable_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0);
+  /*--- Variable export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0);
+  /*--- Function export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0);
+  /*--- Type init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0);
+  /*--- Type import code ---*/
+  __pyx_t_1 = PyImport_ImportModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 9, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_ptype_7cpython_4type_type = __Pyx_ImportType(__pyx_t_1, __Pyx_BUILTIN_MODULE_NAME, "type", 
+  #if defined(PYPY_VERSION_NUM) && PYPY_VERSION_NUM < 0x050B0000
+  sizeof(PyTypeObject),
+  #else
+  sizeof(PyHeapTypeObject),
+  #endif
+  __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_7cpython_4type_type) __PYX_ERR(2, 9, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = PyImport_ImportModule("numpy"); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 206, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_ptype_5numpy_dtype = __Pyx_ImportType(__pyx_t_1, "numpy", "dtype", sizeof(PyArray_Descr), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_dtype) __PYX_ERR(1, 206, __pyx_L1_error)
+  __pyx_ptype_5numpy_flatiter = __Pyx_ImportType(__pyx_t_1, "numpy", "flatiter", sizeof(PyArrayIterObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_flatiter) __PYX_ERR(1, 229, __pyx_L1_error)
+  __pyx_ptype_5numpy_broadcast = __Pyx_ImportType(__pyx_t_1, "numpy", "broadcast", sizeof(PyArrayMultiIterObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_broadcast) __PYX_ERR(1, 233, __pyx_L1_error)
+  __pyx_ptype_5numpy_ndarray = __Pyx_ImportType(__pyx_t_1, "numpy", "ndarray", sizeof(PyArrayObject), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_ndarray) __PYX_ERR(1, 242, __pyx_L1_error)
+  __pyx_ptype_5numpy_ufunc = __Pyx_ImportType(__pyx_t_1, "numpy", "ufunc", sizeof(PyUFuncObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_ufunc) __PYX_ERR(1, 918, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static int __Pyx_modinit_variable_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0);
+  /*--- Variable import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0);
+  /*--- Function import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+
+#if PY_MAJOR_VERSION < 3
+#ifdef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC void
+#else
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#endif
+#else
+#ifdef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC PyObject *
+#else
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#endif
+#endif
+
+
+#if PY_MAJOR_VERSION < 3
+__Pyx_PyMODINIT_FUNC initsoft_nms_cpu(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC initsoft_nms_cpu(void)
+#else
+__Pyx_PyMODINIT_FUNC PyInit_soft_nms_cpu(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC PyInit_soft_nms_cpu(void)
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+{
+  return PyModuleDef_Init(&__pyx_moduledef);
+}
+static CYTHON_SMALL_CODE int __Pyx_check_single_interpreter(void) {
+    #if PY_VERSION_HEX >= 0x030700A1
+    static PY_INT64_T main_interpreter_id = -1;
+    PY_INT64_T current_id = PyInterpreterState_GetID(PyThreadState_Get()->interp);
+    if (main_interpreter_id == -1) {
+        main_interpreter_id = current_id;
+        return (unlikely(current_id == -1)) ? -1 : 0;
+    } else if (unlikely(main_interpreter_id != current_id))
+    #else
+    static PyInterpreterState *main_interpreter = NULL;
+    PyInterpreterState *current_interpreter = PyThreadState_Get()->interp;
+    if (!main_interpreter) {
+        main_interpreter = current_interpreter;
+    } else if (unlikely(main_interpreter != current_interpreter))
+    #endif
+    {
+        PyErr_SetString(
+            PyExc_ImportError,
+            "Interpreter change detected - this module can only be loaded into one interpreter per process.");
+        return -1;
+    }
+    return 0;
+}
+static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name, int allow_none) {
+    PyObject *value = PyObject_GetAttrString(spec, from_name);
+    int result = 0;
+    if (likely(value)) {
+        if (allow_none || value != Py_None) {
+            result = PyDict_SetItemString(moddict, to_name, value);
+        }
+        Py_DECREF(value);
+    } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) {
+        PyErr_Clear();
+    } else {
+        result = -1;
+    }
+    return result;
+}
+static CYTHON_SMALL_CODE PyObject* __pyx_pymod_create(PyObject *spec, CYTHON_UNUSED PyModuleDef *def) {
+    PyObject *module = NULL, *moddict, *modname;
+    if (__Pyx_check_single_interpreter())
+        return NULL;
+    if (__pyx_m)
+        return __Pyx_NewRef(__pyx_m);
+    modname = PyObject_GetAttrString(spec, "name");
+    if (unlikely(!modname)) goto bad;
+    module = PyModule_NewObject(modname);
+    Py_DECREF(modname);
+    if (unlikely(!module)) goto bad;
+    moddict = PyModule_GetDict(module);
+    if (unlikely(!moddict)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__", 0) < 0)) goto bad;
+    return module;
+bad:
+    Py_XDECREF(module);
+    return NULL;
+}
+
+
+static CYTHON_SMALL_CODE int __pyx_pymod_exec_soft_nms_cpu(PyObject *__pyx_pyinit_module)
+#endif
+#endif
+{
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannyDeclarations
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  if (__pyx_m) {
+    if (__pyx_m == __pyx_pyinit_module) return 0;
+    PyErr_SetString(PyExc_RuntimeError, "Module 'soft_nms_cpu' has already been imported. Re-initialisation is not supported.");
+    return -1;
+  }
+  #elif PY_MAJOR_VERSION >= 3
+  if (__pyx_m) return __Pyx_NewRef(__pyx_m);
+  #endif
+  #if CYTHON_REFNANNY
+__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny");
+if (!__Pyx_RefNanny) {
+  PyErr_Clear();
+  __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny");
+  if (!__Pyx_RefNanny)
+      Py_FatalError("failed to import 'refnanny' module");
+}
+#endif
+  __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_soft_nms_cpu(void)", 0);
+  if (__Pyx_check_binary_version() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pxy_PyFrame_Initialize_Offsets
+  __Pxy_PyFrame_Initialize_Offsets();
+  #endif
+  __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pyx_CyFunction_USED
+  if (__pyx_CyFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_FusedFunction_USED
+  if (__pyx_FusedFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Coroutine_USED
+  if (__pyx_Coroutine_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Generator_USED
+  if (__pyx_Generator_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_AsyncGen_USED
+  if (__pyx_AsyncGen_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_StopAsyncIteration_USED
+  if (__pyx_StopAsyncIteration_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  /*--- Library function declarations ---*/
+  /*--- Threads initialization code ---*/
+  #if defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS
+  #ifdef WITH_THREAD /* Python build with threading support? */
+  PyEval_InitThreads();
+  #endif
+  #endif
+  /*--- Module creation code ---*/
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  __pyx_m = __pyx_pyinit_module;
+  Py_INCREF(__pyx_m);
+  #else
+  #if PY_MAJOR_VERSION < 3
+  __pyx_m = Py_InitModule4("soft_nms_cpu", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m);
+  #else
+  __pyx_m = PyModule_Create(&__pyx_moduledef);
+  #endif
+  if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_d);
+  __pyx_b = PyImport_AddModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_b);
+  __pyx_cython_runtime = PyImport_AddModule((char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_cython_runtime);
+  if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  /*--- Initialize various global constants etc. ---*/
+  if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT)
+  if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  if (__pyx_module_is_main_mmdet__ops__nms__soft_nms_cpu) {
+    if (PyObject_SetAttr(__pyx_m, __pyx_n_s_name, __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  }
+  #if PY_MAJOR_VERSION >= 3
+  {
+    PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error)
+    if (!PyDict_GetItemString(modules, "mmdet.ops.nms.soft_nms_cpu")) {
+      if (unlikely(PyDict_SetItemString(modules, "mmdet.ops.nms.soft_nms_cpu", __pyx_m) < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+    }
+  }
+  #endif
+  /*--- Builtin init code ---*/
+  if (__Pyx_InitCachedBuiltins() < 0) goto __pyx_L1_error;
+  /*--- Constants init code ---*/
+  if (__Pyx_InitCachedConstants() < 0) goto __pyx_L1_error;
+  /*--- Global type/function init code ---*/
+  (void)__Pyx_modinit_global_init_code();
+  (void)__Pyx_modinit_variable_export_code();
+  (void)__Pyx_modinit_function_export_code();
+  (void)__Pyx_modinit_type_init_code();
+  if (unlikely(__Pyx_modinit_type_import_code() != 0)) goto __pyx_L1_error;
+  (void)__Pyx_modinit_variable_import_code();
+  (void)__Pyx_modinit_function_import_code();
+  /*--- Execution code ---*/
+  #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED)
+  if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":11
+ * # cython: language_level=3, boundscheck=False
+ * 
+ * import numpy as np             # <<<<<<<<<<<<<<
+ * cimport numpy as np
+ * 
+ */
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_numpy, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_np, __pyx_t_1) < 0) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":22
+ * 
+ * 
+ * def soft_nms_cpu(             # <<<<<<<<<<<<<<
+ *     np.ndarray[float, ndim=2] boxes_in,
+ *     float iou_thr,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_5mmdet_3ops_3nms_12soft_nms_cpu_1soft_nms_cpu, NULL, __pyx_n_s_mmdet_ops_nms_soft_nms_cpu); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 22, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_soft_nms_cpu, __pyx_t_1) < 0) __PYX_ERR(0, 22, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mmdet/ops/nms/src/soft_nms_cpu.pyx":1
+ * # ----------------------------------------------------------             # <<<<<<<<<<<<<<
+ * # Soft-NMS: Improving Object Detection With One Line of Code
+ * # Copyright (c) University of Maryland, College Park
+ */
+  __pyx_t_1 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "../../../../../../home/vasgaoweithu/anaconda3/lib/python3.7/site-packages/Cython/Includes/numpy/__init__.pxd":1046
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /*--- Wrapped vars code ---*/
+
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  if (__pyx_m) {
+    if (__pyx_d) {
+      __Pyx_AddTraceback("init mmdet.ops.nms.soft_nms_cpu", __pyx_clineno, __pyx_lineno, __pyx_filename);
+    }
+    Py_CLEAR(__pyx_m);
+  } else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_ImportError, "init mmdet.ops.nms.soft_nms_cpu");
+  }
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  return (__pyx_m != NULL) ? 0 : -1;
+  #elif PY_MAJOR_VERSION >= 3
+  return __pyx_m;
+  #else
+  return;
+  #endif
+}
+
+/* --- Runtime support code --- */
+/* Refnanny */
+#if CYTHON_REFNANNY
+static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) {
+    PyObject *m = NULL, *p = NULL;
+    void *r = NULL;
+    m = PyImport_ImportModule(modname);
+    if (!m) goto end;
+    p = PyObject_GetAttrString(m, "RefNannyAPI");
+    if (!p) goto end;
+    r = PyLong_AsVoidPtr(p);
+end:
+    Py_XDECREF(p);
+    Py_XDECREF(m);
+    return (__Pyx_RefNannyAPIStruct *)r;
+}
+#endif
+
+/* PyObjectGetAttrStr */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) {
+    PyTypeObject* tp = Py_TYPE(obj);
+    if (likely(tp->tp_getattro))
+        return tp->tp_getattro(obj, attr_name);
+#if PY_MAJOR_VERSION < 3
+    if (likely(tp->tp_getattr))
+        return tp->tp_getattr(obj, PyString_AS_STRING(attr_name));
+#endif
+    return PyObject_GetAttr(obj, attr_name);
+}
+#endif
+
+/* GetBuiltinName */
+static PyObject *__Pyx_GetBuiltinName(PyObject *name) {
+    PyObject* result = __Pyx_PyObject_GetAttrStr(__pyx_b, name);
+    if (unlikely(!result)) {
+        PyErr_Format(PyExc_NameError,
+#if PY_MAJOR_VERSION >= 3
+            "name '%U' is not defined", name);
+#else
+            "name '%.200s' is not defined", PyString_AS_STRING(name));
+#endif
+    }
+    return result;
+}
+
+/* RaiseArgTupleInvalid */
+static void __Pyx_RaiseArgtupleInvalid(
+    const char* func_name,
+    int exact,
+    Py_ssize_t num_min,
+    Py_ssize_t num_max,
+    Py_ssize_t num_found)
+{
+    Py_ssize_t num_expected;
+    const char *more_or_less;
+    if (num_found < num_min) {
+        num_expected = num_min;
+        more_or_less = "at least";
+    } else {
+        num_expected = num_max;
+        more_or_less = "at most";
+    }
+    if (exact) {
+        more_or_less = "exactly";
+    }
+    PyErr_Format(PyExc_TypeError,
+                 "%.200s() takes %.8s %" CYTHON_FORMAT_SSIZE_T "d positional argument%.1s (%" CYTHON_FORMAT_SSIZE_T "d given)",
+                 func_name, more_or_less, num_expected,
+                 (num_expected == 1) ? "" : "s", num_found);
+}
+
+/* RaiseDoubleKeywords */
+static void __Pyx_RaiseDoubleKeywordsError(
+    const char* func_name,
+    PyObject* kw_name)
+{
+    PyErr_Format(PyExc_TypeError,
+        #if PY_MAJOR_VERSION >= 3
+        "%s() got multiple values for keyword argument '%U'", func_name, kw_name);
+        #else
+        "%s() got multiple values for keyword argument '%s'", func_name,
+        PyString_AsString(kw_name));
+        #endif
+}
+
+/* ParseKeywords */
+static int __Pyx_ParseOptionalKeywords(
+    PyObject *kwds,
+    PyObject **argnames[],
+    PyObject *kwds2,
+    PyObject *values[],
+    Py_ssize_t num_pos_args,
+    const char* function_name)
+{
+    PyObject *key = 0, *value = 0;
+    Py_ssize_t pos = 0;
+    PyObject*** name;
+    PyObject*** first_kw_arg = argnames + num_pos_args;
+    while (PyDict_Next(kwds, &pos, &key, &value)) {
+        name = first_kw_arg;
+        while (*name && (**name != key)) name++;
+        if (*name) {
+            values[name-argnames] = value;
+            continue;
+        }
+        name = first_kw_arg;
+        #if PY_MAJOR_VERSION < 3
+        if (likely(PyString_CheckExact(key)) || likely(PyString_Check(key))) {
+            while (*name) {
+                if ((CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**name) == PyString_GET_SIZE(key))
+                        && _PyString_Eq(**name, key)) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    if ((**argname == key) || (
+                            (CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**argname) == PyString_GET_SIZE(key))
+                             && _PyString_Eq(**argname, key))) {
+                        goto arg_passed_twice;
+                    }
+                    argname++;
+                }
+            }
+        } else
+        #endif
+        if (likely(PyUnicode_Check(key))) {
+            while (*name) {
+                int cmp = (**name == key) ? 0 :
+                #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                    (PyUnicode_GET_SIZE(**name) != PyUnicode_GET_SIZE(key)) ? 1 :
+                #endif
+                    PyUnicode_Compare(**name, key);
+                if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                if (cmp == 0) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    int cmp = (**argname == key) ? 0 :
+                    #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                        (PyUnicode_GET_SIZE(**argname) != PyUnicode_GET_SIZE(key)) ? 1 :
+                    #endif
+                        PyUnicode_Compare(**argname, key);
+                    if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                    if (cmp == 0) goto arg_passed_twice;
+                    argname++;
+                }
+            }
+        } else
+            goto invalid_keyword_type;
+        if (kwds2) {
+            if (unlikely(PyDict_SetItem(kwds2, key, value))) goto bad;
+        } else {
+            goto invalid_keyword;
+        }
+    }
+    return 0;
+arg_passed_twice:
+    __Pyx_RaiseDoubleKeywordsError(function_name, key);
+    goto bad;
+invalid_keyword_type:
+    PyErr_Format(PyExc_TypeError,
+        "%.200s() keywords must be strings", function_name);
+    goto bad;
+invalid_keyword:
+    PyErr_Format(PyExc_TypeError,
+    #if PY_MAJOR_VERSION < 3
+        "%.200s() got an unexpected keyword argument '%.200s'",
+        function_name, PyString_AsString(key));
+    #else
+        "%s() got an unexpected keyword argument '%U'",
+        function_name, key);
+    #endif
+bad:
+    return -1;
+}
+
+/* ArgTypeTest */
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact)
+{
+    if (unlikely(!type)) {
+        PyErr_SetString(PyExc_SystemError, "Missing type object");
+        return 0;
+    }
+    else if (exact) {
+        #if PY_MAJOR_VERSION == 2
+        if ((type == &PyBaseString_Type) && likely(__Pyx_PyBaseString_CheckExact(obj))) return 1;
+        #endif
+    }
+    else {
+        if (likely(__Pyx_TypeCheck(obj, type))) return 1;
+    }
+    PyErr_Format(PyExc_TypeError,
+        "Argument '%.200s' has incorrect type (expected %.200s, got %.200s)",
+        name, type->tp_name, Py_TYPE(obj)->tp_name);
+    return 0;
+}
+
+/* IsLittleEndian */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void)
+{
+  union {
+    uint32_t u32;
+    uint8_t u8[4];
+  } S;
+  S.u32 = 0x01020304;
+  return S.u8[0] == 4;
+}
+
+/* BufferFormatCheck */
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type) {
+  stack[0].field = &ctx->root;
+  stack[0].parent_offset = 0;
+  ctx->root.type = type;
+  ctx->root.name = "buffer dtype";
+  ctx->root.offset = 0;
+  ctx->head = stack;
+  ctx->head->field = &ctx->root;
+  ctx->fmt_offset = 0;
+  ctx->head->parent_offset = 0;
+  ctx->new_packmode = '@';
+  ctx->enc_packmode = '@';
+  ctx->new_count = 1;
+  ctx->enc_count = 0;
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  ctx->is_valid_array = 0;
+  ctx->struct_alignment = 0;
+  while (type->typegroup == 'S') {
+    ++ctx->head;
+    ctx->head->field = type->fields;
+    ctx->head->parent_offset = 0;
+    type = type->fields->type;
+  }
+}
+static int __Pyx_BufFmt_ParseNumber(const char** ts) {
+    int count;
+    const char* t = *ts;
+    if (*t < '0' || *t > '9') {
+      return -1;
+    } else {
+        count = *t++ - '0';
+        while (*t >= '0' && *t <= '9') {
+            count *= 10;
+            count += *t++ - '0';
+        }
+    }
+    *ts = t;
+    return count;
+}
+static int __Pyx_BufFmt_ExpectNumber(const char **ts) {
+    int number = __Pyx_BufFmt_ParseNumber(ts);
+    if (number == -1)
+        PyErr_Format(PyExc_ValueError,\
+                     "Does not understand character buffer dtype format string ('%c')", **ts);
+    return number;
+}
+static void __Pyx_BufFmt_RaiseUnexpectedChar(char ch) {
+  PyErr_Format(PyExc_ValueError,
+               "Unexpected format string character: '%c'", ch);
+}
+static const char* __Pyx_BufFmt_DescribeTypeChar(char ch, int is_complex) {
+  switch (ch) {
+    case 'c': return "'char'";
+    case 'b': return "'signed char'";
+    case 'B': return "'unsigned char'";
+    case 'h': return "'short'";
+    case 'H': return "'unsigned short'";
+    case 'i': return "'int'";
+    case 'I': return "'unsigned int'";
+    case 'l': return "'long'";
+    case 'L': return "'unsigned long'";
+    case 'q': return "'long long'";
+    case 'Q': return "'unsigned long long'";
+    case 'f': return (is_complex ? "'complex float'" : "'float'");
+    case 'd': return (is_complex ? "'complex double'" : "'double'");
+    case 'g': return (is_complex ? "'complex long double'" : "'long double'");
+    case 'T': return "a struct";
+    case 'O': return "Python object";
+    case 'P': return "a pointer";
+    case 's': case 'p': return "a string";
+    case 0: return "end";
+    default: return "unparseable format string";
+  }
+}
+static size_t __Pyx_BufFmt_TypeCharToStandardSize(char ch, int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return 2;
+    case 'i': case 'I': case 'l': case 'L': return 4;
+    case 'q': case 'Q': return 8;
+    case 'f': return (is_complex ? 8 : 4);
+    case 'd': return (is_complex ? 16 : 8);
+    case 'g': {
+      PyErr_SetString(PyExc_ValueError, "Python does not define a standard format string size for long double ('g')..");
+      return 0;
+    }
+    case 'O': case 'P': return sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static size_t __Pyx_BufFmt_TypeCharToNativeSize(char ch, int is_complex) {
+  switch (ch) {
+    case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(short);
+    case 'i': case 'I': return sizeof(int);
+    case 'l': case 'L': return sizeof(long);
+    #ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(PY_LONG_LONG);
+    #endif
+    case 'f': return sizeof(float) * (is_complex ? 2 : 1);
+    case 'd': return sizeof(double) * (is_complex ? 2 : 1);
+    case 'g': return sizeof(long double) * (is_complex ? 2 : 1);
+    case 'O': case 'P': return sizeof(void*);
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+typedef struct { char c; short x; } __Pyx_st_short;
+typedef struct { char c; int x; } __Pyx_st_int;
+typedef struct { char c; long x; } __Pyx_st_long;
+typedef struct { char c; float x; } __Pyx_st_float;
+typedef struct { char c; double x; } __Pyx_st_double;
+typedef struct { char c; long double x; } __Pyx_st_longdouble;
+typedef struct { char c; void *x; } __Pyx_st_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { char c; PY_LONG_LONG x; } __Pyx_st_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToAlignment(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_st_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_st_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_st_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_st_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_st_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_st_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_st_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_st_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+/* These are for computing the padding at the end of the struct to align
+   on the first member of the struct. This will probably the same as above,
+   but we don't have any guarantees.
+ */
+typedef struct { short x; char c; } __Pyx_pad_short;
+typedef struct { int x; char c; } __Pyx_pad_int;
+typedef struct { long x; char c; } __Pyx_pad_long;
+typedef struct { float x; char c; } __Pyx_pad_float;
+typedef struct { double x; char c; } __Pyx_pad_double;
+typedef struct { long double x; char c; } __Pyx_pad_longdouble;
+typedef struct { void *x; char c; } __Pyx_pad_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { PY_LONG_LONG x; char c; } __Pyx_pad_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToPadding(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_pad_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_pad_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_pad_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_pad_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_pad_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_pad_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_pad_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_pad_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static char __Pyx_BufFmt_TypeCharToGroup(char ch, int is_complex) {
+  switch (ch) {
+    case 'c':
+        return 'H';
+    case 'b': case 'h': case 'i':
+    case 'l': case 'q': case 's': case 'p':
+        return 'I';
+    case 'B': case 'H': case 'I': case 'L': case 'Q':
+        return 'U';
+    case 'f': case 'd': case 'g':
+        return (is_complex ? 'C' : 'R');
+    case 'O':
+        return 'O';
+    case 'P':
+        return 'P';
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+static void __Pyx_BufFmt_RaiseExpected(__Pyx_BufFmt_Context* ctx) {
+  if (ctx->head == NULL || ctx->head->field == &ctx->root) {
+    const char* expected;
+    const char* quote;
+    if (ctx->head == NULL) {
+      expected = "end";
+      quote = "";
+    } else {
+      expected = ctx->head->field->type->name;
+      quote = "'";
+    }
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected %s%s%s but got %s",
+                 quote, expected, quote,
+                 __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex));
+  } else {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_StructField* parent = (ctx->head - 1)->field;
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected '%s' but got %s in '%s.%s'",
+                 field->type->name, __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex),
+                 parent->type->name, field->name);
+  }
+}
+static int __Pyx_BufFmt_ProcessTypeChunk(__Pyx_BufFmt_Context* ctx) {
+  char group;
+  size_t size, offset, arraysize = 1;
+  if (ctx->enc_type == 0) return 0;
+  if (ctx->head->field->type->arraysize[0]) {
+    int i, ndim = 0;
+    if (ctx->enc_type == 's' || ctx->enc_type == 'p') {
+        ctx->is_valid_array = ctx->head->field->type->ndim == 1;
+        ndim = 1;
+        if (ctx->enc_count != ctx->head->field->type->arraysize[0]) {
+            PyErr_Format(PyExc_ValueError,
+                         "Expected a dimension of size %zu, got %zu",
+                         ctx->head->field->type->arraysize[0], ctx->enc_count);
+            return -1;
+        }
+    }
+    if (!ctx->is_valid_array) {
+      PyErr_Format(PyExc_ValueError, "Expected %d dimensions, got %d",
+                   ctx->head->field->type->ndim, ndim);
+      return -1;
+    }
+    for (i = 0; i < ctx->head->field->type->ndim; i++) {
+      arraysize *= ctx->head->field->type->arraysize[i];
+    }
+    ctx->is_valid_array = 0;
+    ctx->enc_count = 1;
+  }
+  group = __Pyx_BufFmt_TypeCharToGroup(ctx->enc_type, ctx->is_complex);
+  do {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_TypeInfo* type = field->type;
+    if (ctx->enc_packmode == '@' || ctx->enc_packmode == '^') {
+      size = __Pyx_BufFmt_TypeCharToNativeSize(ctx->enc_type, ctx->is_complex);
+    } else {
+      size = __Pyx_BufFmt_TypeCharToStandardSize(ctx->enc_type, ctx->is_complex);
+    }
+    if (ctx->enc_packmode == '@') {
+      size_t align_at = __Pyx_BufFmt_TypeCharToAlignment(ctx->enc_type, ctx->is_complex);
+      size_t align_mod_offset;
+      if (align_at == 0) return -1;
+      align_mod_offset = ctx->fmt_offset % align_at;
+      if (align_mod_offset > 0) ctx->fmt_offset += align_at - align_mod_offset;
+      if (ctx->struct_alignment == 0)
+          ctx->struct_alignment = __Pyx_BufFmt_TypeCharToPadding(ctx->enc_type,
+                                                                 ctx->is_complex);
+    }
+    if (type->size != size || type->typegroup != group) {
+      if (type->typegroup == 'C' && type->fields != NULL) {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        ++ctx->head;
+        ctx->head->field = type->fields;
+        ctx->head->parent_offset = parent_offset;
+        continue;
+      }
+      if ((type->typegroup == 'H' || group == 'H') && type->size == size) {
+      } else {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+      }
+    }
+    offset = ctx->head->parent_offset + field->offset;
+    if (ctx->fmt_offset != offset) {
+      PyErr_Format(PyExc_ValueError,
+                   "Buffer dtype mismatch; next field is at offset %" CYTHON_FORMAT_SSIZE_T "d but %" CYTHON_FORMAT_SSIZE_T "d expected",
+                   (Py_ssize_t)ctx->fmt_offset, (Py_ssize_t)offset);
+      return -1;
+    }
+    ctx->fmt_offset += size;
+    if (arraysize)
+      ctx->fmt_offset += (arraysize - 1) * size;
+    --ctx->enc_count;
+    while (1) {
+      if (field == &ctx->root) {
+        ctx->head = NULL;
+        if (ctx->enc_count != 0) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+        }
+        break;
+      }
+      ctx->head->field = ++field;
+      if (field->type == NULL) {
+        --ctx->head;
+        field = ctx->head->field;
+        continue;
+      } else if (field->type->typegroup == 'S') {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        if (field->type->fields->type == NULL) continue;
+        field = field->type->fields;
+        ++ctx->head;
+        ctx->head->field = field;
+        ctx->head->parent_offset = parent_offset;
+        break;
+      } else {
+        break;
+      }
+    }
+  } while (ctx->enc_count);
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  return 0;
+}
+static PyObject *
+__pyx_buffmt_parse_array(__Pyx_BufFmt_Context* ctx, const char** tsp)
+{
+    const char *ts = *tsp;
+    int i = 0, number;
+    int ndim = ctx->head->field->type->ndim;
+;
+    ++ts;
+    if (ctx->new_count != 1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Cannot handle repeated arrays in format string");
+        return NULL;
+    }
+    if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+    while (*ts && *ts != ')') {
+        switch (*ts) {
+            case ' ': case '\f': case '\r': case '\n': case '\t': case '\v':  continue;
+            default:  break;
+        }
+        number = __Pyx_BufFmt_ExpectNumber(&ts);
+        if (number == -1) return NULL;
+        if (i < ndim && (size_t) number != ctx->head->field->type->arraysize[i])
+            return PyErr_Format(PyExc_ValueError,
+                        "Expected a dimension of size %zu, got %d",
+                        ctx->head->field->type->arraysize[i], number);
+        if (*ts != ',' && *ts != ')')
+            return PyErr_Format(PyExc_ValueError,
+                                "Expected a comma in format string, got '%c'", *ts);
+        if (*ts == ',') ts++;
+        i++;
+    }
+    if (i != ndim)
+        return PyErr_Format(PyExc_ValueError, "Expected %d dimension(s), got %d",
+                            ctx->head->field->type->ndim, i);
+    if (!*ts) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Unexpected end of format string, expected ')'");
+        return NULL;
+    }
+    ctx->is_valid_array = 1;
+    ctx->new_count = 1;
+    *tsp = ++ts;
+    return Py_None;
+}
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts) {
+  int got_Z = 0;
+  while (1) {
+    switch(*ts) {
+      case 0:
+        if (ctx->enc_type != 0 && ctx->head == NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        if (ctx->head != NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        return ts;
+      case ' ':
+      case '\r':
+      case '\n':
+        ++ts;
+        break;
+      case '<':
+        if (!__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Little-endian buffer not supported on big-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '>':
+      case '!':
+        if (__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Big-endian buffer not supported on little-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '=':
+      case '@':
+      case '^':
+        ctx->new_packmode = *ts++;
+        break;
+      case 'T':
+        {
+          const char* ts_after_sub;
+          size_t i, struct_count = ctx->new_count;
+          size_t struct_alignment = ctx->struct_alignment;
+          ctx->new_count = 1;
+          ++ts;
+          if (*ts != '{') {
+            PyErr_SetString(PyExc_ValueError, "Buffer acquisition: Expected '{' after 'T'");
+            return NULL;
+          }
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          ctx->enc_count = 0;
+          ctx->struct_alignment = 0;
+          ++ts;
+          ts_after_sub = ts;
+          for (i = 0; i != struct_count; ++i) {
+            ts_after_sub = __Pyx_BufFmt_CheckString(ctx, ts);
+            if (!ts_after_sub) return NULL;
+          }
+          ts = ts_after_sub;
+          if (struct_alignment) ctx->struct_alignment = struct_alignment;
+        }
+        break;
+      case '}':
+        {
+          size_t alignment = ctx->struct_alignment;
+          ++ts;
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          if (alignment && ctx->fmt_offset % alignment) {
+            ctx->fmt_offset += alignment - (ctx->fmt_offset % alignment);
+          }
+        }
+        return ts;
+      case 'x':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->fmt_offset += ctx->new_count;
+        ctx->new_count = 1;
+        ctx->enc_count = 0;
+        ctx->enc_type = 0;
+        ctx->enc_packmode = ctx->new_packmode;
+        ++ts;
+        break;
+      case 'Z':
+        got_Z = 1;
+        ++ts;
+        if (*ts != 'f' && *ts != 'd' && *ts != 'g') {
+          __Pyx_BufFmt_RaiseUnexpectedChar('Z');
+          return NULL;
+        }
+        CYTHON_FALLTHROUGH;
+      case 'c': case 'b': case 'B': case 'h': case 'H': case 'i': case 'I':
+      case 'l': case 'L': case 'q': case 'Q':
+      case 'f': case 'd': case 'g':
+      case 'O': case 'p':
+        if (ctx->enc_type == *ts && got_Z == ctx->is_complex &&
+            ctx->enc_packmode == ctx->new_packmode) {
+          ctx->enc_count += ctx->new_count;
+          ctx->new_count = 1;
+          got_Z = 0;
+          ++ts;
+          break;
+        }
+        CYTHON_FALLTHROUGH;
+      case 's':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->enc_count = ctx->new_count;
+        ctx->enc_packmode = ctx->new_packmode;
+        ctx->enc_type = *ts;
+        ctx->is_complex = got_Z;
+        ++ts;
+        ctx->new_count = 1;
+        got_Z = 0;
+        break;
+      case ':':
+        ++ts;
+        while(*ts != ':') ++ts;
+        ++ts;
+        break;
+      case '(':
+        if (!__pyx_buffmt_parse_array(ctx, &ts)) return NULL;
+        break;
+      default:
+        {
+          int number = __Pyx_BufFmt_ExpectNumber(&ts);
+          if (number == -1) return NULL;
+          ctx->new_count = (size_t)number;
+        }
+    }
+  }
+}
+
+/* BufferGetAndValidate */
+  static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info) {
+  if (unlikely(info->buf == NULL)) return;
+  if (info->suboffsets == __Pyx_minusones) info->suboffsets = NULL;
+  __Pyx_ReleaseBuffer(info);
+}
+static void __Pyx_ZeroBuffer(Py_buffer* buf) {
+  buf->buf = NULL;
+  buf->obj = NULL;
+  buf->strides = __Pyx_zeros;
+  buf->shape = __Pyx_zeros;
+  buf->suboffsets = __Pyx_minusones;
+}
+static int __Pyx__GetBufferAndValidate(
+        Py_buffer* buf, PyObject* obj,  __Pyx_TypeInfo* dtype, int flags,
+        int nd, int cast, __Pyx_BufFmt_StackElem* stack)
+{
+  buf->buf = NULL;
+  if (unlikely(__Pyx_GetBuffer(obj, buf, flags) == -1)) {
+    __Pyx_ZeroBuffer(buf);
+    return -1;
+  }
+  if (unlikely(buf->ndim != nd)) {
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer has wrong number of dimensions (expected %d, got %d)",
+                 nd, buf->ndim);
+    goto fail;
+  }
+  if (!cast) {
+    __Pyx_BufFmt_Context ctx;
+    __Pyx_BufFmt_Init(&ctx, stack, dtype);
+    if (!__Pyx_BufFmt_CheckString(&ctx, buf->format)) goto fail;
+  }
+  if (unlikely((size_t)buf->itemsize != dtype->size)) {
+    PyErr_Format(PyExc_ValueError,
+      "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "d byte%s) does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "d byte%s)",
+      buf->itemsize, (buf->itemsize > 1) ? "s" : "",
+      dtype->name, (Py_ssize_t)dtype->size, (dtype->size > 1) ? "s" : "");
+    goto fail;
+  }
+  if (buf->suboffsets == NULL) buf->suboffsets = __Pyx_minusones;
+  return 0;
+fail:;
+  __Pyx_SafeReleaseBuffer(buf);
+  return -1;
+}
+
+/* PyFunctionFastCall */
+  #if CYTHON_FAST_PYCALL
+static PyObject* __Pyx_PyFunction_FastCallNoKw(PyCodeObject *co, PyObject **args, Py_ssize_t na,
+                                               PyObject *globals) {
+    PyFrameObject *f;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    PyObject **fastlocals;
+    Py_ssize_t i;
+    PyObject *result;
+    assert(globals != NULL);
+    /* XXX Perhaps we should create a specialized
+       PyFrame_New() that doesn't take locals, but does
+       take builtins without sanity checking them.
+       */
+    assert(tstate != NULL);
+    f = PyFrame_New(tstate, co, globals, NULL);
+    if (f == NULL) {
+        return NULL;
+    }
+    fastlocals = __Pyx_PyFrame_GetLocalsplus(f);
+    for (i = 0; i < na; i++) {
+        Py_INCREF(*args);
+        fastlocals[i] = *args++;
+    }
+    result = PyEval_EvalFrameEx(f,0);
+    ++tstate->recursion_depth;
+    Py_DECREF(f);
+    --tstate->recursion_depth;
+    return result;
+}
+#if 1 || PY_VERSION_HEX < 0x030600B1
+static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs) {
+    PyCodeObject *co = (PyCodeObject *)PyFunction_GET_CODE(func);
+    PyObject *globals = PyFunction_GET_GLOBALS(func);
+    PyObject *argdefs = PyFunction_GET_DEFAULTS(func);
+    PyObject *closure;
+#if PY_MAJOR_VERSION >= 3
+    PyObject *kwdefs;
+#endif
+    PyObject *kwtuple, **k;
+    PyObject **d;
+    Py_ssize_t nd;
+    Py_ssize_t nk;
+    PyObject *result;
+    assert(kwargs == NULL || PyDict_Check(kwargs));
+    nk = kwargs ? PyDict_Size(kwargs) : 0;
+    if (Py_EnterRecursiveCall((char*)" while calling a Python object")) {
+        return NULL;
+    }
+    if (
+#if PY_MAJOR_VERSION >= 3
+            co->co_kwonlyargcount == 0 &&
+#endif
+            likely(kwargs == NULL || nk == 0) &&
+            co->co_flags == (CO_OPTIMIZED | CO_NEWLOCALS | CO_NOFREE)) {
+        if (argdefs == NULL && co->co_argcount == nargs) {
+            result = __Pyx_PyFunction_FastCallNoKw(co, args, nargs, globals);
+            goto done;
+        }
+        else if (nargs == 0 && argdefs != NULL
+                 && co->co_argcount == Py_SIZE(argdefs)) {
+            /* function called with no arguments, but all parameters have
+               a default value: use default values as arguments .*/
+            args = &PyTuple_GET_ITEM(argdefs, 0);
+            result =__Pyx_PyFunction_FastCallNoKw(co, args, Py_SIZE(argdefs), globals);
+            goto done;
+        }
+    }
+    if (kwargs != NULL) {
+        Py_ssize_t pos, i;
+        kwtuple = PyTuple_New(2 * nk);
+        if (kwtuple == NULL) {
+            result = NULL;
+            goto done;
+        }
+        k = &PyTuple_GET_ITEM(kwtuple, 0);
+        pos = i = 0;
+        while (PyDict_Next(kwargs, &pos, &k[i], &k[i+1])) {
+            Py_INCREF(k[i]);
+            Py_INCREF(k[i+1]);
+            i += 2;
+        }
+        nk = i / 2;
+    }
+    else {
+        kwtuple = NULL;
+        k = NULL;
+    }
+    closure = PyFunction_GET_CLOSURE(func);
+#if PY_MAJOR_VERSION >= 3
+    kwdefs = PyFunction_GET_KW_DEFAULTS(func);
+#endif
+    if (argdefs != NULL) {
+        d = &PyTuple_GET_ITEM(argdefs, 0);
+        nd = Py_SIZE(argdefs);
+    }
+    else {
+        d = NULL;
+        nd = 0;
+    }
+#if PY_MAJOR_VERSION >= 3
+    result = PyEval_EvalCodeEx((PyObject*)co, globals, (PyObject *)NULL,
+                               args, (int)nargs,
+                               k, (int)nk,
+                               d, (int)nd, kwdefs, closure);
+#else
+    result = PyEval_EvalCodeEx(co, globals, (PyObject *)NULL,
+                               args, (int)nargs,
+                               k, (int)nk,
+                               d, (int)nd, closure);
+#endif
+    Py_XDECREF(kwtuple);
+done:
+    Py_LeaveRecursiveCall();
+    return result;
+}
+#endif
+#endif
+
+/* PyObjectCall */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw) {
+    PyObject *result;
+    ternaryfunc call = func->ob_type->tp_call;
+    if (unlikely(!call))
+        return PyObject_Call(func, arg, kw);
+    if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object")))
+        return NULL;
+    result = (*call)(func, arg, kw);
+    Py_LeaveRecursiveCall();
+    if (unlikely(!result) && unlikely(!PyErr_Occurred())) {
+        PyErr_SetString(
+            PyExc_SystemError,
+            "NULL result without error in PyObject_Call");
+    }
+    return result;
+}
+#endif
+
+/* PyObjectCallMethO */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg) {
+    PyObject *self, *result;
+    PyCFunction cfunc;
+    cfunc = PyCFunction_GET_FUNCTION(func);
+    self = PyCFunction_GET_SELF(func);
+    if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object")))
+        return NULL;
+    result = cfunc(self, arg);
+    Py_LeaveRecursiveCall();
+    if (unlikely(!result) && unlikely(!PyErr_Occurred())) {
+        PyErr_SetString(
+            PyExc_SystemError,
+            "NULL result without error in PyObject_Call");
+    }
+    return result;
+}
+#endif
+
+/* PyObjectCallNoArg */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func) {
+#if CYTHON_FAST_PYCALL
+    if (PyFunction_Check(func)) {
+        return __Pyx_PyFunction_FastCall(func, NULL, 0);
+    }
+#endif
+#ifdef __Pyx_CyFunction_USED
+    if (likely(PyCFunction_Check(func) || __Pyx_CyFunction_Check(func)))
+#else
+    if (likely(PyCFunction_Check(func)))
+#endif
+    {
+        if (likely(PyCFunction_GET_FLAGS(func) & METH_NOARGS)) {
+            return __Pyx_PyObject_CallMethO(func, NULL);
+        }
+    }
+    return __Pyx_PyObject_Call(func, __pyx_empty_tuple, NULL);
+}
+#endif
+
+/* PyCFunctionFastCall */
+  #if CYTHON_FAST_PYCCALL
+static CYTHON_INLINE PyObject * __Pyx_PyCFunction_FastCall(PyObject *func_obj, PyObject **args, Py_ssize_t nargs) {
+    PyCFunctionObject *func = (PyCFunctionObject*)func_obj;
+    PyCFunction meth = PyCFunction_GET_FUNCTION(func);
+    PyObject *self = PyCFunction_GET_SELF(func);
+    int flags = PyCFunction_GET_FLAGS(func);
+    assert(PyCFunction_Check(func));
+    assert(METH_FASTCALL == (flags & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS | METH_STACKLESS)));
+    assert(nargs >= 0);
+    assert(nargs == 0 || args != NULL);
+    /* _PyCFunction_FastCallDict() must not be called with an exception set,
+       because it may clear it (directly or indirectly) and so the
+       caller loses its exception */
+    assert(!PyErr_Occurred());
+    if ((PY_VERSION_HEX < 0x030700A0) || unlikely(flags & METH_KEYWORDS)) {
+        return (*((__Pyx_PyCFunctionFastWithKeywords)(void*)meth)) (self, args, nargs, NULL);
+    } else {
+        return (*((__Pyx_PyCFunctionFast)(void*)meth)) (self, args, nargs);
+    }
+}
+#endif
+
+/* PyObjectCallOneArg */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static PyObject* __Pyx__PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+    PyObject *result;
+    PyObject *args = PyTuple_New(1);
+    if (unlikely(!args)) return NULL;
+    Py_INCREF(arg);
+    PyTuple_SET_ITEM(args, 0, arg);
+    result = __Pyx_PyObject_Call(func, args, NULL);
+    Py_DECREF(args);
+    return result;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+#if CYTHON_FAST_PYCALL
+    if (PyFunction_Check(func)) {
+        return __Pyx_PyFunction_FastCall(func, &arg, 1);
+    }
+#endif
+    if (likely(PyCFunction_Check(func))) {
+        if (likely(PyCFunction_GET_FLAGS(func) & METH_O)) {
+            return __Pyx_PyObject_CallMethO(func, arg);
+#if CYTHON_FAST_PYCCALL
+        } else if (PyCFunction_GET_FLAGS(func) & METH_FASTCALL) {
+            return __Pyx_PyCFunction_FastCall(func, &arg, 1);
+#endif
+        }
+    }
+    return __Pyx__PyObject_CallOneArg(func, arg);
+}
+#else
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+    PyObject *result;
+    PyObject *args = PyTuple_Pack(1, arg);
+    if (unlikely(!args)) return NULL;
+    result = __Pyx_PyObject_Call(func, args, NULL);
+    Py_DECREF(args);
+    return result;
+}
+#endif
+
+/* GetItemInt */
+  static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j) {
+    PyObject *r;
+    if (!j) return NULL;
+    r = PyObject_GetItem(o, j);
+    Py_DECREF(j);
+    return r;
+}
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i,
+                                                              CYTHON_NCP_UNUSED int wraparound,
+                                                              CYTHON_NCP_UNUSED int boundscheck) {
+#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+    Py_ssize_t wrapped_i = i;
+    if (wraparound & unlikely(i < 0)) {
+        wrapped_i += PyList_GET_SIZE(o);
+    }
+    if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyList_GET_SIZE(o)))) {
+        PyObject *r = PyList_GET_ITEM(o, wrapped_i);
+        Py_INCREF(r);
+        return r;
+    }
+    return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i));
+#else
+    return PySequence_GetItem(o, i);
+#endif
+}
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i,
+                                                              CYTHON_NCP_UNUSED int wraparound,
+                                                              CYTHON_NCP_UNUSED int boundscheck) {
+#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+    Py_ssize_t wrapped_i = i;
+    if (wraparound & unlikely(i < 0)) {
+        wrapped_i += PyTuple_GET_SIZE(o);
+    }
+    if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyTuple_GET_SIZE(o)))) {
+        PyObject *r = PyTuple_GET_ITEM(o, wrapped_i);
+        Py_INCREF(r);
+        return r;
+    }
+    return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i));
+#else
+    return PySequence_GetItem(o, i);
+#endif
+}
+static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i, int is_list,
+                                                     CYTHON_NCP_UNUSED int wraparound,
+                                                     CYTHON_NCP_UNUSED int boundscheck) {
+#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS
+    if (is_list || PyList_CheckExact(o)) {
+        Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyList_GET_SIZE(o);
+        if ((!boundscheck) || (likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o))))) {
+            PyObject *r = PyList_GET_ITEM(o, n);
+            Py_INCREF(r);
+            return r;
+        }
+    }
+    else if (PyTuple_CheckExact(o)) {
+        Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyTuple_GET_SIZE(o);
+        if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyTuple_GET_SIZE(o)))) {
+            PyObject *r = PyTuple_GET_ITEM(o, n);
+            Py_INCREF(r);
+            return r;
+        }
+    } else {
+        PySequenceMethods *m = Py_TYPE(o)->tp_as_sequence;
+        if (likely(m && m->sq_item)) {
+            if (wraparound && unlikely(i < 0) && likely(m->sq_length)) {
+                Py_ssize_t l = m->sq_length(o);
+                if (likely(l >= 0)) {
+                    i += l;
+                } else {
+                    if (!PyErr_ExceptionMatches(PyExc_OverflowError))
+                        return NULL;
+                    PyErr_Clear();
+                }
+            }
+            return m->sq_item(o, i);
+        }
+    }
+#else
+    if (is_list || PySequence_Check(o)) {
+        return PySequence_GetItem(o, i);
+    }
+#endif
+    return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i));
+}
+
+/* PyDictVersioning */
+  #if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    return likely(dict) ? __PYX_GET_DICT_VERSION(dict) : 0;
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj) {
+    PyObject **dictptr = NULL;
+    Py_ssize_t offset = Py_TYPE(obj)->tp_dictoffset;
+    if (offset) {
+#if CYTHON_COMPILING_IN_CPYTHON
+        dictptr = (likely(offset > 0)) ? (PyObject **) ((char *)obj + offset) : _PyObject_GetDictPtr(obj);
+#else
+        dictptr = _PyObject_GetDictPtr(obj);
+#endif
+    }
+    return (dictptr && *dictptr) ? __PYX_GET_DICT_VERSION(*dictptr) : 0;
+}
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    if (unlikely(!dict) || unlikely(tp_dict_version != __PYX_GET_DICT_VERSION(dict)))
+        return 0;
+    return obj_dict_version == __Pyx_get_object_dict_version(obj);
+}
+#endif
+
+/* GetModuleGlobalName */
+  #if CYTHON_USE_DICT_VERSIONS
+static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value)
+#else
+static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name)
+#endif
+{
+    PyObject *result;
+#if !CYTHON_AVOID_BORROWED_REFS
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1
+    result = _PyDict_GetItem_KnownHash(__pyx_d, name, ((PyASCIIObject *) name)->hash);
+    __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version)
+    if (likely(result)) {
+        return __Pyx_NewRef(result);
+    } else if (unlikely(PyErr_Occurred())) {
+        return NULL;
+    }
+#else
+    result = PyDict_GetItem(__pyx_d, name);
+    __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version)
+    if (likely(result)) {
+        return __Pyx_NewRef(result);
+    }
+#endif
+#else
+    result = PyObject_GetItem(__pyx_d, name);
+    __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version)
+    if (likely(result)) {
+        return __Pyx_NewRef(result);
+    }
+    PyErr_Clear();
+#endif
+    return __Pyx_GetBuiltinName(name);
+}
+
+/* PyObjectCall2Args */
+  static CYTHON_UNUSED PyObject* __Pyx_PyObject_Call2Args(PyObject* function, PyObject* arg1, PyObject* arg2) {
+    PyObject *args, *result = NULL;
+    #if CYTHON_FAST_PYCALL
+    if (PyFunction_Check(function)) {
+        PyObject *args[2] = {arg1, arg2};
+        return __Pyx_PyFunction_FastCall(function, args, 2);
+    }
+    #endif
+    #if CYTHON_FAST_PYCCALL
+    if (__Pyx_PyFastCFunction_Check(function)) {
+        PyObject *args[2] = {arg1, arg2};
+        return __Pyx_PyCFunction_FastCall(function, args, 2);
+    }
+    #endif
+    args = PyTuple_New(2);
+    if (unlikely(!args)) goto done;
+    Py_INCREF(arg1);
+    PyTuple_SET_ITEM(args, 0, arg1);
+    Py_INCREF(arg2);
+    PyTuple_SET_ITEM(args, 1, arg2);
+    Py_INCREF(function);
+    result = __Pyx_PyObject_Call(function, args, NULL);
+    Py_DECREF(args);
+    Py_DECREF(function);
+done:
+    return result;
+}
+
+/* ObjectGetItem */
+  #if CYTHON_USE_TYPE_SLOTS
+static PyObject *__Pyx_PyObject_GetIndex(PyObject *obj, PyObject* index) {
+    PyObject *runerr;
+    Py_ssize_t key_value;
+    PySequenceMethods *m = Py_TYPE(obj)->tp_as_sequence;
+    if (unlikely(!(m && m->sq_item))) {
+        PyErr_Format(PyExc_TypeError, "'%.200s' object is not subscriptable", Py_TYPE(obj)->tp_name);
+        return NULL;
+    }
+    key_value = __Pyx_PyIndex_AsSsize_t(index);
+    if (likely(key_value != -1 || !(runerr = PyErr_Occurred()))) {
+        return __Pyx_GetItemInt_Fast(obj, key_value, 0, 1, 1);
+    }
+    if (PyErr_GivenExceptionMatches(runerr, PyExc_OverflowError)) {
+        PyErr_Clear();
+        PyErr_Format(PyExc_IndexError, "cannot fit '%.200s' into an index-sized integer", Py_TYPE(index)->tp_name);
+    }
+    return NULL;
+}
+static PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject* key) {
+    PyMappingMethods *m = Py_TYPE(obj)->tp_as_mapping;
+    if (likely(m && m->mp_subscript)) {
+        return m->mp_subscript(obj, key);
+    }
+    return __Pyx_PyObject_GetIndex(obj, key);
+}
+#endif
+
+/* PyIntBinop */
+  #if !CYTHON_COMPILING_IN_PYPY
+static PyObject* __Pyx_PyInt_AddObjC(PyObject *op1, PyObject *op2, CYTHON_UNUSED long intval, int inplace, int zerodivision_check) {
+    (void)inplace;
+    (void)zerodivision_check;
+    #if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_CheckExact(op1))) {
+        const long b = intval;
+        long x;
+        long a = PyInt_AS_LONG(op1);
+            x = (long)((unsigned long)a + b);
+            if (likely((x^a) >= 0 || (x^b) >= 0))
+                return PyInt_FromLong(x);
+            return PyLong_Type.tp_as_number->nb_add(op1, op2);
+    }
+    #endif
+    #if CYTHON_USE_PYLONG_INTERNALS
+    if (likely(PyLong_CheckExact(op1))) {
+        const long b = intval;
+        long a, x;
+#ifdef HAVE_LONG_LONG
+        const PY_LONG_LONG llb = intval;
+        PY_LONG_LONG lla, llx;
+#endif
+        const digit* digits = ((PyLongObject*)op1)->ob_digit;
+        const Py_ssize_t size = Py_SIZE(op1);
+        if (likely(__Pyx_sst_abs(size) <= 1)) {
+            a = likely(size) ? digits[0] : 0;
+            if (size == -1) a = -a;
+        } else {
+            switch (size) {
+                case -2:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        a = -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 2 * PyLong_SHIFT) {
+                        lla = -(PY_LONG_LONG) (((((unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                case 2:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        a = (long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 2 * PyLong_SHIFT) {
+                        lla = (PY_LONG_LONG) (((((unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                case -3:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        a = -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 3 * PyLong_SHIFT) {
+                        lla = -(PY_LONG_LONG) (((((((unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                case 3:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        a = (long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 3 * PyLong_SHIFT) {
+                        lla = (PY_LONG_LONG) (((((((unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                case -4:
+                    if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                        a = -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 4 * PyLong_SHIFT) {
+                        lla = -(PY_LONG_LONG) (((((((((unsigned PY_LONG_LONG)digits[3]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                case 4:
+                    if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                        a = (long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]));
+                        break;
+#ifdef HAVE_LONG_LONG
+                    } else if (8 * sizeof(PY_LONG_LONG) - 1 > 4 * PyLong_SHIFT) {
+                        lla = (PY_LONG_LONG) (((((((((unsigned PY_LONG_LONG)digits[3]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0]));
+                        goto long_long;
+#endif
+                    }
+                    CYTHON_FALLTHROUGH;
+                default: return PyLong_Type.tp_as_number->nb_add(op1, op2);
+            }
+        }
+                x = a + b;
+            return PyLong_FromLong(x);
+#ifdef HAVE_LONG_LONG
+        long_long:
+                llx = lla + llb;
+            return PyLong_FromLongLong(llx);
+#endif
+        
+        
+    }
+    #endif
+    if (PyFloat_CheckExact(op1)) {
+        const long b = intval;
+        double a = PyFloat_AS_DOUBLE(op1);
+            double result;
+            PyFPE_START_PROTECT("add", return NULL)
+            result = ((double)a) + (double)b;
+            PyFPE_END_PROTECT(result)
+            return PyFloat_FromDouble(result);
+    }
+    return (inplace ? PyNumber_InPlaceAdd : PyNumber_Add)(op1, op2);
+}
+#endif
+
+/* SetItemInt */
+  static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v) {
+    int r;
+    if (!j) return -1;
+    r = PyObject_SetItem(o, j, v);
+    Py_DECREF(j);
+    return r;
+}
+static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v, int is_list,
+                                               CYTHON_NCP_UNUSED int wraparound, CYTHON_NCP_UNUSED int boundscheck) {
+#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS
+    if (is_list || PyList_CheckExact(o)) {
+        Py_ssize_t n = (!wraparound) ? i : ((likely(i >= 0)) ? i : i + PyList_GET_SIZE(o));
+        if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o)))) {
+            PyObject* old = PyList_GET_ITEM(o, n);
+            Py_INCREF(v);
+            PyList_SET_ITEM(o, n, v);
+            Py_DECREF(old);
+            return 1;
+        }
+    } else {
+        PySequenceMethods *m = Py_TYPE(o)->tp_as_sequence;
+        if (likely(m && m->sq_ass_item)) {
+            if (wraparound && unlikely(i < 0) && likely(m->sq_length)) {
+                Py_ssize_t l = m->sq_length(o);
+                if (likely(l >= 0)) {
+                    i += l;
+                } else {
+                    if (!PyErr_ExceptionMatches(PyExc_OverflowError))
+                        return -1;
+                    PyErr_Clear();
+                }
+            }
+            return m->sq_ass_item(o, i, v);
+        }
+    }
+#else
+#if CYTHON_COMPILING_IN_PYPY
+    if (is_list || (PySequence_Check(o) && !PyDict_Check(o)))
+#else
+    if (is_list || PySequence_Check(o))
+#endif
+    {
+        return PySequence_SetItem(o, i, v);
+    }
+#endif
+    return __Pyx_SetItemInt_Generic(o, PyInt_FromSsize_t(i), v);
+}
+
+/* SliceObject */
+  static CYTHON_INLINE PyObject* __Pyx_PyObject_GetSlice(PyObject* obj,
+        Py_ssize_t cstart, Py_ssize_t cstop,
+        PyObject** _py_start, PyObject** _py_stop, PyObject** _py_slice,
+        int has_cstart, int has_cstop, CYTHON_UNUSED int wraparound) {
+#if CYTHON_USE_TYPE_SLOTS
+    PyMappingMethods* mp;
+#if PY_MAJOR_VERSION < 3
+    PySequenceMethods* ms = Py_TYPE(obj)->tp_as_sequence;
+    if (likely(ms && ms->sq_slice)) {
+        if (!has_cstart) {
+            if (_py_start && (*_py_start != Py_None)) {
+                cstart = __Pyx_PyIndex_AsSsize_t(*_py_start);
+                if ((cstart == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad;
+            } else
+                cstart = 0;
+        }
+        if (!has_cstop) {
+            if (_py_stop && (*_py_stop != Py_None)) {
+                cstop = __Pyx_PyIndex_AsSsize_t(*_py_stop);
+                if ((cstop == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad;
+            } else
+                cstop = PY_SSIZE_T_MAX;
+        }
+        if (wraparound && unlikely((cstart < 0) | (cstop < 0)) && likely(ms->sq_length)) {
+            Py_ssize_t l = ms->sq_length(obj);
+            if (likely(l >= 0)) {
+                if (cstop < 0) {
+                    cstop += l;
+                    if (cstop < 0) cstop = 0;
+                }
+                if (cstart < 0) {
+                    cstart += l;
+                    if (cstart < 0) cstart = 0;
+                }
+            } else {
+                if (!PyErr_ExceptionMatches(PyExc_OverflowError))
+                    goto bad;
+                PyErr_Clear();
+            }
+        }
+        return ms->sq_slice(obj, cstart, cstop);
+    }
+#endif
+    mp = Py_TYPE(obj)->tp_as_mapping;
+    if (likely(mp && mp->mp_subscript))
+#endif
+    {
+        PyObject* result;
+        PyObject *py_slice, *py_start, *py_stop;
+        if (_py_slice) {
+            py_slice = *_py_slice;
+        } else {
+            PyObject* owned_start = NULL;
+            PyObject* owned_stop = NULL;
+            if (_py_start) {
+                py_start = *_py_start;
+            } else {
+                if (has_cstart) {
+                    owned_start = py_start = PyInt_FromSsize_t(cstart);
+                    if (unlikely(!py_start)) goto bad;
+                } else
+                    py_start = Py_None;
+            }
+            if (_py_stop) {
+                py_stop = *_py_stop;
+            } else {
+                if (has_cstop) {
+                    owned_stop = py_stop = PyInt_FromSsize_t(cstop);
+                    if (unlikely(!py_stop)) {
+                        Py_XDECREF(owned_start);
+                        goto bad;
+                    }
+                } else
+                    py_stop = Py_None;
+            }
+            py_slice = PySlice_New(py_start, py_stop, Py_None);
+            Py_XDECREF(owned_start);
+            Py_XDECREF(owned_stop);
+            if (unlikely(!py_slice)) goto bad;
+        }
+#if CYTHON_USE_TYPE_SLOTS
+        result = mp->mp_subscript(obj, py_slice);
+#else
+        result = PyObject_GetItem(obj, py_slice);
+#endif
+        if (!_py_slice) {
+            Py_DECREF(py_slice);
+        }
+        return result;
+    }
+    PyErr_Format(PyExc_TypeError,
+        "'%.200s' object is unsliceable", Py_TYPE(obj)->tp_name);
+bad:
+    return NULL;
+}
+
+/* PyErrFetchRestore */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    tmp_type = tstate->curexc_type;
+    tmp_value = tstate->curexc_value;
+    tmp_tb = tstate->curexc_traceback;
+    tstate->curexc_type = type;
+    tstate->curexc_value = value;
+    tstate->curexc_traceback = tb;
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    *type = tstate->curexc_type;
+    *value = tstate->curexc_value;
+    *tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+}
+#endif
+
+/* RaiseException */
+  #if PY_MAJOR_VERSION < 3
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb,
+                        CYTHON_UNUSED PyObject *cause) {
+    __Pyx_PyThreadState_declare
+    Py_XINCREF(type);
+    if (!value || value == Py_None)
+        value = NULL;
+    else
+        Py_INCREF(value);
+    if (!tb || tb == Py_None)
+        tb = NULL;
+    else {
+        Py_INCREF(tb);
+        if (!PyTraceBack_Check(tb)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: arg 3 must be a traceback or None");
+            goto raise_error;
+        }
+    }
+    if (PyType_Check(type)) {
+#if CYTHON_COMPILING_IN_PYPY
+        if (!value) {
+            Py_INCREF(Py_None);
+            value = Py_None;
+        }
+#endif
+        PyErr_NormalizeException(&type, &value, &tb);
+    } else {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto raise_error;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(type);
+        Py_INCREF(type);
+        if (!PyType_IsSubtype((PyTypeObject *)type, (PyTypeObject *)PyExc_BaseException)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: exception class must be a subclass of BaseException");
+            goto raise_error;
+        }
+    }
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrRestore(type, value, tb);
+    return;
+raise_error:
+    Py_XDECREF(value);
+    Py_XDECREF(type);
+    Py_XDECREF(tb);
+    return;
+}
+#else
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) {
+    PyObject* owned_instance = NULL;
+    if (tb == Py_None) {
+        tb = 0;
+    } else if (tb && !PyTraceBack_Check(tb)) {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: arg 3 must be a traceback or None");
+        goto bad;
+    }
+    if (value == Py_None)
+        value = 0;
+    if (PyExceptionInstance_Check(type)) {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto bad;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(value);
+    } else if (PyExceptionClass_Check(type)) {
+        PyObject *instance_class = NULL;
+        if (value && PyExceptionInstance_Check(value)) {
+            instance_class = (PyObject*) Py_TYPE(value);
+            if (instance_class != type) {
+                int is_subclass = PyObject_IsSubclass(instance_class, type);
+                if (!is_subclass) {
+                    instance_class = NULL;
+                } else if (unlikely(is_subclass == -1)) {
+                    goto bad;
+                } else {
+                    type = instance_class;
+                }
+            }
+        }
+        if (!instance_class) {
+            PyObject *args;
+            if (!value)
+                args = PyTuple_New(0);
+            else if (PyTuple_Check(value)) {
+                Py_INCREF(value);
+                args = value;
+            } else
+                args = PyTuple_Pack(1, value);
+            if (!args)
+                goto bad;
+            owned_instance = PyObject_Call(type, args, NULL);
+            Py_DECREF(args);
+            if (!owned_instance)
+                goto bad;
+            value = owned_instance;
+            if (!PyExceptionInstance_Check(value)) {
+                PyErr_Format(PyExc_TypeError,
+                             "calling %R should have returned an instance of "
+                             "BaseException, not %R",
+                             type, Py_TYPE(value));
+                goto bad;
+            }
+        }
+    } else {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: exception class must be a subclass of BaseException");
+        goto bad;
+    }
+    if (cause) {
+        PyObject *fixed_cause;
+        if (cause == Py_None) {
+            fixed_cause = NULL;
+        } else if (PyExceptionClass_Check(cause)) {
+            fixed_cause = PyObject_CallObject(cause, NULL);
+            if (fixed_cause == NULL)
+                goto bad;
+        } else if (PyExceptionInstance_Check(cause)) {
+            fixed_cause = cause;
+            Py_INCREF(fixed_cause);
+        } else {
+            PyErr_SetString(PyExc_TypeError,
+                            "exception causes must derive from "
+                            "BaseException");
+            goto bad;
+        }
+        PyException_SetCause(value, fixed_cause);
+    }
+    PyErr_SetObject(type, value);
+    if (tb) {
+#if CYTHON_COMPILING_IN_PYPY
+        PyObject *tmp_type, *tmp_value, *tmp_tb;
+        PyErr_Fetch(&tmp_type, &tmp_value, &tmp_tb);
+        Py_INCREF(tb);
+        PyErr_Restore(tmp_type, tmp_value, tb);
+        Py_XDECREF(tmp_tb);
+#else
+        PyThreadState *tstate = __Pyx_PyThreadState_Current;
+        PyObject* tmp_tb = tstate->curexc_traceback;
+        if (tb != tmp_tb) {
+            Py_INCREF(tb);
+            tstate->curexc_traceback = tb;
+            Py_XDECREF(tmp_tb);
+        }
+#endif
+    }
+bad:
+    Py_XDECREF(owned_instance);
+    return;
+}
+#endif
+
+/* DictGetItem */
+  #if PY_MAJOR_VERSION >= 3 && !CYTHON_COMPILING_IN_PYPY
+static PyObject *__Pyx_PyDict_GetItem(PyObject *d, PyObject* key) {
+    PyObject *value;
+    value = PyDict_GetItemWithError(d, key);
+    if (unlikely(!value)) {
+        if (!PyErr_Occurred()) {
+            if (unlikely(PyTuple_Check(key))) {
+                PyObject* args = PyTuple_Pack(1, key);
+                if (likely(args)) {
+                    PyErr_SetObject(PyExc_KeyError, args);
+                    Py_DECREF(args);
+                }
+            } else {
+                PyErr_SetObject(PyExc_KeyError, key);
+            }
+        }
+        return NULL;
+    }
+    Py_INCREF(value);
+    return value;
+}
+#endif
+
+/* RaiseTooManyValuesToUnpack */
+  static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected) {
+    PyErr_Format(PyExc_ValueError,
+                 "too many values to unpack (expected %" CYTHON_FORMAT_SSIZE_T "d)", expected);
+}
+
+/* RaiseNeedMoreValuesToUnpack */
+  static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index) {
+    PyErr_Format(PyExc_ValueError,
+                 "need more than %" CYTHON_FORMAT_SSIZE_T "d value%.1s to unpack",
+                 index, (index == 1) ? "" : "s");
+}
+
+/* RaiseNoneIterError */
+  static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void) {
+    PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable");
+}
+
+/* ExtTypeTest */
+  static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type) {
+    if (unlikely(!type)) {
+        PyErr_SetString(PyExc_SystemError, "Missing type object");
+        return 0;
+    }
+    if (likely(__Pyx_TypeCheck(obj, type)))
+        return 1;
+    PyErr_Format(PyExc_TypeError, "Cannot convert %.200s to %.200s",
+                 Py_TYPE(obj)->tp_name, type->tp_name);
+    return 0;
+}
+
+/* GetTopmostException */
+  #if CYTHON_USE_EXC_INFO_STACK
+static _PyErr_StackItem *
+__Pyx_PyErr_GetTopmostException(PyThreadState *tstate)
+{
+    _PyErr_StackItem *exc_info = tstate->exc_info;
+    while ((exc_info->exc_type == NULL || exc_info->exc_type == Py_None) &&
+           exc_info->previous_item != NULL)
+    {
+        exc_info = exc_info->previous_item;
+    }
+    return exc_info;
+}
+#endif
+
+/* SaveResetException */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    #if CYTHON_USE_EXC_INFO_STACK
+    _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate);
+    *type = exc_info->exc_type;
+    *value = exc_info->exc_value;
+    *tb = exc_info->exc_traceback;
+    #else
+    *type = tstate->exc_type;
+    *value = tstate->exc_value;
+    *tb = tstate->exc_traceback;
+    #endif
+    Py_XINCREF(*type);
+    Py_XINCREF(*value);
+    Py_XINCREF(*tb);
+}
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    #if CYTHON_USE_EXC_INFO_STACK
+    _PyErr_StackItem *exc_info = tstate->exc_info;
+    tmp_type = exc_info->exc_type;
+    tmp_value = exc_info->exc_value;
+    tmp_tb = exc_info->exc_traceback;
+    exc_info->exc_type = type;
+    exc_info->exc_value = value;
+    exc_info->exc_traceback = tb;
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = type;
+    tstate->exc_value = value;
+    tstate->exc_traceback = tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+#endif
+
+/* PyErrExceptionMatches */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx_PyErr_ExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        if (__Pyx_PyErr_GivenExceptionMatches(exc_type, PyTuple_GET_ITEM(tuple, i))) return 1;
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err) {
+    PyObject *exc_type = tstate->curexc_type;
+    if (exc_type == err) return 1;
+    if (unlikely(!exc_type)) return 0;
+    if (unlikely(PyTuple_Check(err)))
+        return __Pyx_PyErr_ExceptionMatchesTuple(exc_type, err);
+    return __Pyx_PyErr_GivenExceptionMatches(exc_type, err);
+}
+#endif
+
+/* GetException */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb)
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb)
+#endif
+{
+    PyObject *local_type, *local_value, *local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    local_type = tstate->curexc_type;
+    local_value = tstate->curexc_value;
+    local_tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+#else
+    PyErr_Fetch(&local_type, &local_value, &local_tb);
+#endif
+    PyErr_NormalizeException(&local_type, &local_value, &local_tb);
+#if CYTHON_FAST_THREAD_STATE
+    if (unlikely(tstate->curexc_type))
+#else
+    if (unlikely(PyErr_Occurred()))
+#endif
+        goto bad;
+    #if PY_MAJOR_VERSION >= 3
+    if (local_tb) {
+        if (unlikely(PyException_SetTraceback(local_value, local_tb) < 0))
+            goto bad;
+    }
+    #endif
+    Py_XINCREF(local_tb);
+    Py_XINCREF(local_type);
+    Py_XINCREF(local_value);
+    *type = local_type;
+    *value = local_value;
+    *tb = local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    #if CYTHON_USE_EXC_INFO_STACK
+    {
+        _PyErr_StackItem *exc_info = tstate->exc_info;
+        tmp_type = exc_info->exc_type;
+        tmp_value = exc_info->exc_value;
+        tmp_tb = exc_info->exc_traceback;
+        exc_info->exc_type = local_type;
+        exc_info->exc_value = local_value;
+        exc_info->exc_traceback = local_tb;
+    }
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = local_type;
+    tstate->exc_value = local_value;
+    tstate->exc_traceback = local_tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+#else
+    PyErr_SetExcInfo(local_type, local_value, local_tb);
+#endif
+    return 0;
+bad:
+    *type = 0;
+    *value = 0;
+    *tb = 0;
+    Py_XDECREF(local_type);
+    Py_XDECREF(local_value);
+    Py_XDECREF(local_tb);
+    return -1;
+}
+
+/* TypeImport */
+  #ifndef __PYX_HAVE_RT_ImportType
+#define __PYX_HAVE_RT_ImportType
+static PyTypeObject *__Pyx_ImportType(PyObject *module, const char *module_name, const char *class_name,
+    size_t size, enum __Pyx_ImportType_CheckSize check_size)
+{
+    PyObject *result = 0;
+    char warning[200];
+    Py_ssize_t basicsize;
+#ifdef Py_LIMITED_API
+    PyObject *py_basicsize;
+#endif
+    result = PyObject_GetAttrString(module, class_name);
+    if (!result)
+        goto bad;
+    if (!PyType_Check(result)) {
+        PyErr_Format(PyExc_TypeError,
+            "%.200s.%.200s is not a type object",
+            module_name, class_name);
+        goto bad;
+    }
+#ifndef Py_LIMITED_API
+    basicsize = ((PyTypeObject *)result)->tp_basicsize;
+#else
+    py_basicsize = PyObject_GetAttrString(result, "__basicsize__");
+    if (!py_basicsize)
+        goto bad;
+    basicsize = PyLong_AsSsize_t(py_basicsize);
+    Py_DECREF(py_basicsize);
+    py_basicsize = 0;
+    if (basicsize == (Py_ssize_t)-1 && PyErr_Occurred())
+        goto bad;
+#endif
+    if ((size_t)basicsize < size) {
+        PyErr_Format(PyExc_ValueError,
+            "%.200s.%.200s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        goto bad;
+    }
+    if (check_size == __Pyx_ImportType_CheckSize_Error && (size_t)basicsize != size) {
+        PyErr_Format(PyExc_ValueError,
+            "%.200s.%.200s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        goto bad;
+    }
+    else if (check_size == __Pyx_ImportType_CheckSize_Warn && (size_t)basicsize > size) {
+        PyOS_snprintf(warning, sizeof(warning),
+            "%s.%s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        if (PyErr_WarnEx(NULL, warning, 0) < 0) goto bad;
+    }
+    return (PyTypeObject *)result;
+bad:
+    Py_XDECREF(result);
+    return NULL;
+}
+#endif
+
+/* Import */
+  static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) {
+    PyObject *empty_list = 0;
+    PyObject *module = 0;
+    PyObject *global_dict = 0;
+    PyObject *empty_dict = 0;
+    PyObject *list;
+    #if PY_MAJOR_VERSION < 3
+    PyObject *py_import;
+    py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import);
+    if (!py_import)
+        goto bad;
+    #endif
+    if (from_list)
+        list = from_list;
+    else {
+        empty_list = PyList_New(0);
+        if (!empty_list)
+            goto bad;
+        list = empty_list;
+    }
+    global_dict = PyModule_GetDict(__pyx_m);
+    if (!global_dict)
+        goto bad;
+    empty_dict = PyDict_New();
+    if (!empty_dict)
+        goto bad;
+    {
+        #if PY_MAJOR_VERSION >= 3
+        if (level == -1) {
+            if (strchr(__Pyx_MODULE_NAME, '.')) {
+                module = PyImport_ImportModuleLevelObject(
+                    name, global_dict, empty_dict, list, 1);
+                if (!module) {
+                    if (!PyErr_ExceptionMatches(PyExc_ImportError))
+                        goto bad;
+                    PyErr_Clear();
+                }
+            }
+            level = 0;
+        }
+        #endif
+        if (!module) {
+            #if PY_MAJOR_VERSION < 3
+            PyObject *py_level = PyInt_FromLong(level);
+            if (!py_level)
+                goto bad;
+            module = PyObject_CallFunctionObjArgs(py_import,
+                name, global_dict, empty_dict, list, py_level, (PyObject *)NULL);
+            Py_DECREF(py_level);
+            #else
+            module = PyImport_ImportModuleLevelObject(
+                name, global_dict, empty_dict, list, level);
+            #endif
+        }
+    }
+bad:
+    #if PY_MAJOR_VERSION < 3
+    Py_XDECREF(py_import);
+    #endif
+    Py_XDECREF(empty_list);
+    Py_XDECREF(empty_dict);
+    return module;
+}
+
+/* CLineInTraceback */
+  #ifndef CYTHON_CLINE_IN_TRACEBACK
+static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line) {
+    PyObject *use_cline;
+    PyObject *ptype, *pvalue, *ptraceback;
+#if CYTHON_COMPILING_IN_CPYTHON
+    PyObject **cython_runtime_dict;
+#endif
+    if (unlikely(!__pyx_cython_runtime)) {
+        return c_line;
+    }
+    __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback);
+#if CYTHON_COMPILING_IN_CPYTHON
+    cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime);
+    if (likely(cython_runtime_dict)) {
+        __PYX_PY_DICT_LOOKUP_IF_MODIFIED(
+            use_cline, *cython_runtime_dict,
+            __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback))
+    } else
+#endif
+    {
+      PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback);
+      if (use_cline_obj) {
+        use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True;
+        Py_DECREF(use_cline_obj);
+      } else {
+        PyErr_Clear();
+        use_cline = NULL;
+      }
+    }
+    if (!use_cline) {
+        c_line = 0;
+        PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False);
+    }
+    else if (use_cline == Py_False || (use_cline != Py_True && PyObject_Not(use_cline) != 0)) {
+        c_line = 0;
+    }
+    __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback);
+    return c_line;
+}
+#endif
+
+/* CodeObjectCache */
+  static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) {
+    int start = 0, mid = 0, end = count - 1;
+    if (end >= 0 && code_line > entries[end].code_line) {
+        return count;
+    }
+    while (start < end) {
+        mid = start + (end - start) / 2;
+        if (code_line < entries[mid].code_line) {
+            end = mid;
+        } else if (code_line > entries[mid].code_line) {
+             start = mid + 1;
+        } else {
+            return mid;
+        }
+    }
+    if (code_line <= entries[mid].code_line) {
+        return mid;
+    } else {
+        return mid + 1;
+    }
+}
+static PyCodeObject *__pyx_find_code_object(int code_line) {
+    PyCodeObject* code_object;
+    int pos;
+    if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) {
+        return NULL;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) {
+        return NULL;
+    }
+    code_object = __pyx_code_cache.entries[pos].code_object;
+    Py_INCREF(code_object);
+    return code_object;
+}
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) {
+    int pos, i;
+    __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries;
+    if (unlikely(!code_line)) {
+        return;
+    }
+    if (unlikely(!entries)) {
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (likely(entries)) {
+            __pyx_code_cache.entries = entries;
+            __pyx_code_cache.max_count = 64;
+            __pyx_code_cache.count = 1;
+            entries[0].code_line = code_line;
+            entries[0].code_object = code_object;
+            Py_INCREF(code_object);
+        }
+        return;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) {
+        PyCodeObject* tmp = entries[pos].code_object;
+        entries[pos].code_object = code_object;
+        Py_DECREF(tmp);
+        return;
+    }
+    if (__pyx_code_cache.count == __pyx_code_cache.max_count) {
+        int new_max = __pyx_code_cache.max_count + 64;
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc(
+            __pyx_code_cache.entries, (size_t)new_max*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (unlikely(!entries)) {
+            return;
+        }
+        __pyx_code_cache.entries = entries;
+        __pyx_code_cache.max_count = new_max;
+    }
+    for (i=__pyx_code_cache.count; i>pos; i--) {
+        entries[i] = entries[i-1];
+    }
+    entries[pos].code_line = code_line;
+    entries[pos].code_object = code_object;
+    __pyx_code_cache.count++;
+    Py_INCREF(code_object);
+}
+
+/* AddTraceback */
+  #include "compile.h"
+#include "frameobject.h"
+#include "traceback.h"
+static PyCodeObject* __Pyx_CreateCodeObjectForTraceback(
+            const char *funcname, int c_line,
+            int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyObject *py_srcfile = 0;
+    PyObject *py_funcname = 0;
+    #if PY_MAJOR_VERSION < 3
+    py_srcfile = PyString_FromString(filename);
+    #else
+    py_srcfile = PyUnicode_FromString(filename);
+    #endif
+    if (!py_srcfile) goto bad;
+    if (c_line) {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #else
+        py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #endif
+    }
+    else {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromString(funcname);
+        #else
+        py_funcname = PyUnicode_FromString(funcname);
+        #endif
+    }
+    if (!py_funcname) goto bad;
+    py_code = __Pyx_PyCode_New(
+        0,
+        0,
+        0,
+        0,
+        0,
+        __pyx_empty_bytes, /*PyObject *code,*/
+        __pyx_empty_tuple, /*PyObject *consts,*/
+        __pyx_empty_tuple, /*PyObject *names,*/
+        __pyx_empty_tuple, /*PyObject *varnames,*/
+        __pyx_empty_tuple, /*PyObject *freevars,*/
+        __pyx_empty_tuple, /*PyObject *cellvars,*/
+        py_srcfile,   /*PyObject *filename,*/
+        py_funcname,  /*PyObject *name,*/
+        py_line,
+        __pyx_empty_bytes  /*PyObject *lnotab*/
+    );
+    Py_DECREF(py_srcfile);
+    Py_DECREF(py_funcname);
+    return py_code;
+bad:
+    Py_XDECREF(py_srcfile);
+    Py_XDECREF(py_funcname);
+    return NULL;
+}
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyFrameObject *py_frame = 0;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    if (c_line) {
+        c_line = __Pyx_CLineForTraceback(tstate, c_line);
+    }
+    py_code = __pyx_find_code_object(c_line ? -c_line : py_line);
+    if (!py_code) {
+        py_code = __Pyx_CreateCodeObjectForTraceback(
+            funcname, c_line, py_line, filename);
+        if (!py_code) goto bad;
+        __pyx_insert_code_object(c_line ? -c_line : py_line, py_code);
+    }
+    py_frame = PyFrame_New(
+        tstate,            /*PyThreadState *tstate,*/
+        py_code,           /*PyCodeObject *code,*/
+        __pyx_d,    /*PyObject *globals,*/
+        0                  /*PyObject *locals*/
+    );
+    if (!py_frame) goto bad;
+    __Pyx_PyFrame_SetLineNumber(py_frame, py_line);
+    PyTraceBack_Here(py_frame);
+bad:
+    Py_XDECREF(py_code);
+    Py_XDECREF(py_frame);
+}
+
+#if PY_MAJOR_VERSION < 3
+static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags) {
+    if (PyObject_CheckBuffer(obj)) return PyObject_GetBuffer(obj, view, flags);
+        if (__Pyx_TypeCheck(obj, __pyx_ptype_5numpy_ndarray)) return __pyx_pw_5numpy_7ndarray_1__getbuffer__(obj, view, flags);
+    PyErr_Format(PyExc_TypeError, "'%.200s' does not have the buffer interface", Py_TYPE(obj)->tp_name);
+    return -1;
+}
+static void __Pyx_ReleaseBuffer(Py_buffer *view) {
+    PyObject *obj = view->obj;
+    if (!obj) return;
+    if (PyObject_CheckBuffer(obj)) {
+        PyBuffer_Release(view);
+        return;
+    }
+    if ((0)) {}
+        else if (__Pyx_TypeCheck(obj, __pyx_ptype_5numpy_ndarray)) __pyx_pw_5numpy_7ndarray_3__releasebuffer__(obj, view);
+    view->obj = NULL;
+    Py_DECREF(obj);
+}
+#endif
+
+
+  /* CIntFromPyVerify */
+  #define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0)
+#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1)
+#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\
+    {\
+        func_type value = func_value;\
+        if (sizeof(target_type) < sizeof(func_type)) {\
+            if (unlikely(value != (func_type) (target_type) value)) {\
+                func_type zero = 0;\
+                if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\
+                    return (target_type) -1;\
+                if (is_unsigned && unlikely(value < zero))\
+                    goto raise_neg_overflow;\
+                else\
+                    goto raise_overflow;\
+            }\
+        }\
+        return (target_type) value;\
+    }
+
+/* CIntToPy */
+  static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) {
+    const long neg_one = (long) ((long) 0 - (long) 1), const_zero = (long) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(long) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(long) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(long) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(long),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntToPy */
+  static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value) {
+    const int neg_one = (int) ((int) 0 - (int) 1), const_zero = (int) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(int) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(int) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(int) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(int),
+                                     little, !is_unsigned);
+    }
+}
+
+/* Declarations */
+  #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return ::std::complex< float >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return x + y*(__pyx_t_float_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      __pyx_t_float_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+  #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabsf(b.real) >= fabsf(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                float r = b.imag / b.real;
+                float s = (float)(1.0) / (b.real + b.imag * r);
+                return __pyx_t_float_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            float r = b.real / b.imag;
+            float s = (float)(1.0) / (b.imag + b.real * r);
+            return __pyx_t_float_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            float denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_float_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrtf(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypotf(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+            __pyx_t_float_complex z;
+            float r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    float denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = powf(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2f(0.0, -1.0);
+                }
+            } else {
+                r = __Pyx_c_abs_float(a);
+                theta = atan2f(a.imag, a.real);
+            }
+            lnr = logf(r);
+            z_r = expf(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cosf(z_theta);
+            z.imag = z_r * sinf(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* Declarations */
+  #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return ::std::complex< double >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return x + y*(__pyx_t_double_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      __pyx_t_double_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+  #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabs(b.real) >= fabs(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                double r = b.imag / b.real;
+                double s = (double)(1.0) / (b.real + b.imag * r);
+                return __pyx_t_double_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            double r = b.real / b.imag;
+            double s = (double)(1.0) / (b.imag + b.real * r);
+            return __pyx_t_double_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            double denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_double_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrt(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypot(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+            __pyx_t_double_complex z;
+            double r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    double denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = pow(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2(0.0, -1.0);
+                }
+            } else {
+                r = __Pyx_c_abs_double(a);
+                theta = atan2(a.imag, a.real);
+            }
+            lnr = log(r);
+            z_r = exp(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cos(z_theta);
+            z.imag = z_r * sin(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* CIntToPy */
+  static CYTHON_INLINE PyObject* __Pyx_PyInt_From_enum__NPY_TYPES(enum NPY_TYPES value) {
+    const enum NPY_TYPES neg_one = (enum NPY_TYPES) ((enum NPY_TYPES) 0 - (enum NPY_TYPES) 1), const_zero = (enum NPY_TYPES) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(enum NPY_TYPES) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(enum NPY_TYPES) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(enum NPY_TYPES),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntFromPy */
+  static CYTHON_INLINE unsigned int __Pyx_PyInt_As_unsigned_int(PyObject *x) {
+    const unsigned int neg_one = (unsigned int) ((unsigned int) 0 - (unsigned int) 1), const_zero = (unsigned int) 0;
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(unsigned int) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(unsigned int, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (unsigned int) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (unsigned int) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(unsigned int, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(unsigned int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) >= 2 * PyLong_SHIFT) {
+                            return (unsigned int) (((((unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(unsigned int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) >= 3 * PyLong_SHIFT) {
+                            return (unsigned int) (((((((unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(unsigned int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) >= 4 * PyLong_SHIFT) {
+                            return (unsigned int) (((((((((unsigned int)digits[3]) << PyLong_SHIFT) | (unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (unsigned int) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(unsigned int) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(unsigned int, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(unsigned int) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(unsigned int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (unsigned int) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(unsigned int, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(unsigned int,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(unsigned int) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 2 * PyLong_SHIFT) {
+                            return (unsigned int) (((unsigned int)-1)*(((((unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(unsigned int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 2 * PyLong_SHIFT) {
+                            return (unsigned int) ((((((unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(unsigned int) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 3 * PyLong_SHIFT) {
+                            return (unsigned int) (((unsigned int)-1)*(((((((unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(unsigned int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 3 * PyLong_SHIFT) {
+                            return (unsigned int) ((((((((unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(unsigned int) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 4 * PyLong_SHIFT) {
+                            return (unsigned int) (((unsigned int)-1)*(((((((((unsigned int)digits[3]) << PyLong_SHIFT) | (unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(unsigned int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(unsigned int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(unsigned int) - 1 > 4 * PyLong_SHIFT) {
+                            return (unsigned int) ((((((((((unsigned int)digits[3]) << PyLong_SHIFT) | (unsigned int)digits[2]) << PyLong_SHIFT) | (unsigned int)digits[1]) << PyLong_SHIFT) | (unsigned int)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(unsigned int) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(unsigned int, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(unsigned int) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(unsigned int, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            unsigned int val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (unsigned int) -1;
+        }
+    } else {
+        unsigned int val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (unsigned int) -1;
+        val = __Pyx_PyInt_As_unsigned_int(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to unsigned int");
+    return (unsigned int) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to unsigned int");
+    return (unsigned int) -1;
+}
+
+/* CIntFromPy */
+  static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) {
+    const int neg_one = (int) ((int) 0 - (int) 1), const_zero = (int) 0;
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(int) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (int) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(int, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 2 * PyLong_SHIFT) {
+                            return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 3 * PyLong_SHIFT) {
+                            return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 4 * PyLong_SHIFT) {
+                            return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (int) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(int) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(int, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(int,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(int) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(int) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            int val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (int) -1;
+        }
+    } else {
+        int val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (int) -1;
+        val = __Pyx_PyInt_As_int(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to int");
+    return (int) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to int");
+    return (int) -1;
+}
+
+/* CIntFromPy */
+  static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) {
+    const long neg_one = (long) ((long) 0 - (long) 1), const_zero = (long) 0;
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(long) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (long) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(long, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 2 * PyLong_SHIFT) {
+                            return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 3 * PyLong_SHIFT) {
+                            return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 4 * PyLong_SHIFT) {
+                            return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (long) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(long) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(long, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(long,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(long) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(long) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            long val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (long) -1;
+        }
+    } else {
+        long val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (long) -1;
+        val = __Pyx_PyInt_As_long(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to long");
+    return (long) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to long");
+    return (long) -1;
+}
+
+/* FastTypeChecks */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) {
+    while (a) {
+        a = a->tp_base;
+        if (a == b)
+            return 1;
+    }
+    return b == &PyBaseObject_Type;
+}
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) {
+    PyObject *mro;
+    if (a == b) return 1;
+    mro = a->tp_mro;
+    if (likely(mro)) {
+        Py_ssize_t i, n;
+        n = PyTuple_GET_SIZE(mro);
+        for (i = 0; i < n; i++) {
+            if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b)
+                return 1;
+        }
+        return 0;
+    }
+    return __Pyx_InBases(a, b);
+}
+#if PY_MAJOR_VERSION == 2
+static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) {
+    PyObject *exception, *value, *tb;
+    int res;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&exception, &value, &tb);
+    res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0;
+    if (unlikely(res == -1)) {
+        PyErr_WriteUnraisable(err);
+        res = 0;
+    }
+    if (!res) {
+        res = PyObject_IsSubclass(err, exc_type2);
+        if (unlikely(res == -1)) {
+            PyErr_WriteUnraisable(err);
+            res = 0;
+        }
+    }
+    __Pyx_ErrRestore(exception, value, tb);
+    return res;
+}
+#else
+static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) {
+    int res = exc_type1 ? __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type1) : 0;
+    if (!res) {
+        res = __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2);
+    }
+    return res;
+}
+#endif
+static int __Pyx_PyErr_GivenExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    assert(PyExceptionClass_Check(exc_type));
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        PyObject *t = PyTuple_GET_ITEM(tuple, i);
+        #if PY_MAJOR_VERSION < 3
+        if (likely(exc_type == t)) return 1;
+        #endif
+        if (likely(PyExceptionClass_Check(t))) {
+            if (__Pyx_inner_PyErr_GivenExceptionMatches2(exc_type, NULL, t)) return 1;
+        } else {
+        }
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject* exc_type) {
+    if (likely(err == exc_type)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        if (likely(PyExceptionClass_Check(exc_type))) {
+            return __Pyx_inner_PyErr_GivenExceptionMatches2(err, NULL, exc_type);
+        } else if (likely(PyTuple_Check(exc_type))) {
+            return __Pyx_PyErr_GivenExceptionMatchesTuple(err, exc_type);
+        } else {
+        }
+    }
+    return PyErr_GivenExceptionMatches(err, exc_type);
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *exc_type1, PyObject *exc_type2) {
+    assert(PyExceptionClass_Check(exc_type1));
+    assert(PyExceptionClass_Check(exc_type2));
+    if (likely(err == exc_type1 || err == exc_type2)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        return __Pyx_inner_PyErr_GivenExceptionMatches2(err, exc_type1, exc_type2);
+    }
+    return (PyErr_GivenExceptionMatches(err, exc_type1) || PyErr_GivenExceptionMatches(err, exc_type2));
+}
+#endif
+
+/* CheckBinaryVersion */
+  static int __Pyx_check_binary_version(void) {
+    char ctversion[4], rtversion[4];
+    PyOS_snprintf(ctversion, 4, "%d.%d", PY_MAJOR_VERSION, PY_MINOR_VERSION);
+    PyOS_snprintf(rtversion, 4, "%s", Py_GetVersion());
+    if (ctversion[0] != rtversion[0] || ctversion[2] != rtversion[2]) {
+        char message[200];
+        PyOS_snprintf(message, sizeof(message),
+                      "compiletime version %s of module '%.100s' "
+                      "does not match runtime version %s",
+                      ctversion, __Pyx_MODULE_NAME, rtversion);
+        return PyErr_WarnEx(NULL, message, 1);
+    }
+    return 0;
+}
+
+/* InitStrings */
+  static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) {
+    while (t->p) {
+        #if PY_MAJOR_VERSION < 3
+        if (t->is_unicode) {
+            *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL);
+        } else if (t->intern) {
+            *t->p = PyString_InternFromString(t->s);
+        } else {
+            *t->p = PyString_FromStringAndSize(t->s, t->n - 1);
+        }
+        #else
+        if (t->is_unicode | t->is_str) {
+            if (t->intern) {
+                *t->p = PyUnicode_InternFromString(t->s);
+            } else if (t->encoding) {
+                *t->p = PyUnicode_Decode(t->s, t->n - 1, t->encoding, NULL);
+            } else {
+                *t->p = PyUnicode_FromStringAndSize(t->s, t->n - 1);
+            }
+        } else {
+            *t->p = PyBytes_FromStringAndSize(t->s, t->n - 1);
+        }
+        #endif
+        if (!*t->p)
+            return -1;
+        if (PyObject_Hash(*t->p) == -1)
+            return -1;
+        ++t;
+    }
+    return 0;
+}
+
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) {
+    return __Pyx_PyUnicode_FromStringAndSize(c_str, (Py_ssize_t)strlen(c_str));
+}
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) {
+    Py_ssize_t ignore;
+    return __Pyx_PyObject_AsStringAndSize(o, &ignore);
+}
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+#if !CYTHON_PEP393_ENABLED
+static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    char* defenc_c;
+    PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL);
+    if (!defenc) return NULL;
+    defenc_c = PyBytes_AS_STRING(defenc);
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    {
+        char* end = defenc_c + PyBytes_GET_SIZE(defenc);
+        char* c;
+        for (c = defenc_c; c < end; c++) {
+            if ((unsigned char) (*c) >= 128) {
+                PyUnicode_AsASCIIString(o);
+                return NULL;
+            }
+        }
+    }
+#endif
+    *length = PyBytes_GET_SIZE(defenc);
+    return defenc_c;
+}
+#else
+static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL;
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    if (likely(PyUnicode_IS_ASCII(o))) {
+        *length = PyUnicode_GET_LENGTH(o);
+        return PyUnicode_AsUTF8(o);
+    } else {
+        PyUnicode_AsASCIIString(o);
+        return NULL;
+    }
+#else
+    return PyUnicode_AsUTF8AndSize(o, length);
+#endif
+}
+#endif
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+    if (
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+            __Pyx_sys_getdefaultencoding_not_ascii &&
+#endif
+            PyUnicode_Check(o)) {
+        return __Pyx_PyUnicode_AsStringAndSize(o, length);
+    } else
+#endif
+#if (!CYTHON_COMPILING_IN_PYPY) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE))
+    if (PyByteArray_Check(o)) {
+        *length = PyByteArray_GET_SIZE(o);
+        return PyByteArray_AS_STRING(o);
+    } else
+#endif
+    {
+        char* result;
+        int r = PyBytes_AsStringAndSize(o, &result, length);
+        if (unlikely(r < 0)) {
+            return NULL;
+        } else {
+            return result;
+        }
+    }
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) {
+   int is_true = x == Py_True;
+   if (is_true | (x == Py_False) | (x == Py_None)) return is_true;
+   else return PyObject_IsTrue(x);
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject* x) {
+    int retval;
+    if (unlikely(!x)) return -1;
+    retval = __Pyx_PyObject_IsTrue(x);
+    Py_DECREF(x);
+    return retval;
+}
+static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) {
+#if PY_MAJOR_VERSION >= 3
+    if (PyLong_Check(result)) {
+        if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1,
+                "__int__ returned non-int (type %.200s).  "
+                "The ability to return an instance of a strict subclass of int "
+                "is deprecated, and may be removed in a future version of Python.",
+                Py_TYPE(result)->tp_name)) {
+            Py_DECREF(result);
+            return NULL;
+        }
+        return result;
+    }
+#endif
+    PyErr_Format(PyExc_TypeError,
+                 "__%.4s__ returned non-%.4s (type %.200s)",
+                 type_name, type_name, Py_TYPE(result)->tp_name);
+    Py_DECREF(result);
+    return NULL;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) {
+#if CYTHON_USE_TYPE_SLOTS
+  PyNumberMethods *m;
+#endif
+  const char *name = NULL;
+  PyObject *res = NULL;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_Check(x) || PyLong_Check(x)))
+#else
+  if (likely(PyLong_Check(x)))
+#endif
+    return __Pyx_NewRef(x);
+#if CYTHON_USE_TYPE_SLOTS
+  m = Py_TYPE(x)->tp_as_number;
+  #if PY_MAJOR_VERSION < 3
+  if (m && m->nb_int) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  else if (m && m->nb_long) {
+    name = "long";
+    res = m->nb_long(x);
+  }
+  #else
+  if (likely(m && m->nb_int)) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  #endif
+#else
+  if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) {
+    res = PyNumber_Int(x);
+  }
+#endif
+  if (likely(res)) {
+#if PY_MAJOR_VERSION < 3
+    if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) {
+#else
+    if (unlikely(!PyLong_CheckExact(res))) {
+#endif
+        return __Pyx_PyNumber_IntOrLongWrongResultType(res, name);
+    }
+  }
+  else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_TypeError,
+                    "an integer is required");
+  }
+  return res;
+}
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) {
+  Py_ssize_t ival;
+  PyObject *x;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_CheckExact(b))) {
+    if (sizeof(Py_ssize_t) >= sizeof(long))
+        return PyInt_AS_LONG(b);
+    else
+        return PyInt_AsSsize_t(b);
+  }
+#endif
+  if (likely(PyLong_CheckExact(b))) {
+    #if CYTHON_USE_PYLONG_INTERNALS
+    const digit* digits = ((PyLongObject*)b)->ob_digit;
+    const Py_ssize_t size = Py_SIZE(b);
+    if (likely(__Pyx_sst_abs(size) <= 1)) {
+        ival = likely(size) ? digits[0] : 0;
+        if (size == -1) ival = -ival;
+        return ival;
+    } else {
+      switch (size) {
+         case 2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+      }
+    }
+    #endif
+    return PyLong_AsSsize_t(b);
+  }
+  x = PyNumber_Index(b);
+  if (!x) return -1;
+  ival = PyInt_AsSsize_t(x);
+  Py_DECREF(x);
+  return ival;
+}
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b) {
+  return b ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False);
+}
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) {
+    return PyInt_FromSize_t(ival);
+}
+
+
+#endif /* Py_PYTHON_H */
diff --git a/lib/ops/nms/src/soft_nms_cpu.pyx b/lib/ops/nms/src/soft_nms_cpu.pyx
new file mode 100644
index 0000000..97f53f1
--- /dev/null
+++ b/lib/ops/nms/src/soft_nms_cpu.pyx
@@ -0,0 +1,127 @@
+# ----------------------------------------------------------
+# Soft-NMS: Improving Object Detection With One Line of Code
+# Copyright (c) University of Maryland, College Park
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Navaneeth Bodla and Bharat Singh
+# Modified by Kai Chen
+# ----------------------------------------------------------
+
+# cython: language_level=3, boundscheck=False
+
+import numpy as np
+cimport numpy as np
+
+
+cdef inline np.float32_t max(np.float32_t a, np.float32_t b):
+    return a if a >= b else b
+
+cdef inline np.float32_t min(np.float32_t a, np.float32_t b):
+    return a if a <= b else b
+
+
+def soft_nms_cpu(
+    np.ndarray[float, ndim=2] boxes_in,
+    float iou_thr,
+    unsigned int method=1,
+    float sigma=0.5,
+    float min_score=0.001,
+):
+    boxes = boxes_in.copy()
+    cdef int N = boxes.shape[0]
+    cdef float iw, ih, box_area
+    cdef float ua
+    cdef int pos = 0
+    cdef float maxscore = 0
+    cdef int maxpos = 0
+    cdef float x1, x2, y1, y2, tx1, tx2, ty1, ty2, ts, area, weight, ov
+    inds = np.arange(N)
+
+    for i in range(N):
+        maxscore = boxes[i, 4]
+        maxpos = i
+
+        tx1 = boxes[i, 0]
+        ty1 = boxes[i, 1]
+        tx2 = boxes[i, 2]
+        ty2 = boxes[i, 3]
+        ts = boxes[i, 4]
+        ti = inds[i]
+
+        pos = i + 1
+        # get max box
+        while pos < N:
+            if maxscore < boxes[pos, 4]:
+                maxscore = boxes[pos, 4]
+                maxpos = pos
+            pos = pos + 1
+
+        # add max box as a detection
+        boxes[i, 0] = boxes[maxpos, 0]
+        boxes[i, 1] = boxes[maxpos, 1]
+        boxes[i, 2] = boxes[maxpos, 2]
+        boxes[i, 3] = boxes[maxpos, 3]
+        boxes[i, 4] = boxes[maxpos, 4]
+        inds[i] = inds[maxpos]
+
+        # swap ith box with position of max box
+        boxes[maxpos, 0] = tx1
+        boxes[maxpos, 1] = ty1
+        boxes[maxpos, 2] = tx2
+        boxes[maxpos, 3] = ty2
+        boxes[maxpos, 4] = ts
+        inds[maxpos] = ti
+
+        tx1 = boxes[i, 0]
+        ty1 = boxes[i, 1]
+        tx2 = boxes[i, 2]
+        ty2 = boxes[i, 3]
+        ts = boxes[i, 4]
+
+        pos = i + 1
+        # NMS iterations, note that N changes if detection boxes fall below
+        # threshold
+        while pos < N:
+            x1 = boxes[pos, 0]
+            y1 = boxes[pos, 1]
+            x2 = boxes[pos, 2]
+            y2 = boxes[pos, 3]
+            s = boxes[pos, 4]
+
+            area = (x2 - x1 + 1) * (y2 - y1 + 1)
+            iw = (min(tx2, x2) - max(tx1, x1) + 1)
+            if iw > 0:
+                ih = (min(ty2, y2) - max(ty1, y1) + 1)
+                if ih > 0:
+                    ua = float((tx2 - tx1 + 1) * (ty2 - ty1 + 1) + area - iw * ih)
+                    ov = iw * ih / ua  # iou between max box and detection box
+
+                    if method == 1:  # linear
+                        if ov > iou_thr:
+                            weight = 1 - ov
+                        else:
+                            weight = 1
+                    elif method == 2:  # gaussian
+                        weight = np.exp(-(ov * ov) / sigma)
+                    else:  # original NMS
+                        if ov > iou_thr:
+                            weight = 0
+                        else:
+                            weight = 1
+
+                    boxes[pos, 4] = weight * boxes[pos, 4]
+
+                    # if box score falls below threshold, discard the box by
+                    # swapping with last box update N
+                    if boxes[pos, 4] < min_score:
+                        boxes[pos, 0] = boxes[N-1, 0]
+                        boxes[pos, 1] = boxes[N-1, 1]
+                        boxes[pos, 2] = boxes[N-1, 2]
+                        boxes[pos, 3] = boxes[N-1, 3]
+                        boxes[pos, 4] = boxes[N-1, 4]
+                        inds[pos] = inds[N - 1]
+                        N = N - 1
+                        pos = pos - 1
+
+            pos = pos + 1
+
+    return boxes[:N], inds[:N]
diff --git a/lib/ops/roi_align/__init__.py b/lib/ops/roi_align/__init__.py
new file mode 100644
index 0000000..6da9829
--- /dev/null
+++ b/lib/ops/roi_align/__init__.py
@@ -0,0 +1,3 @@
+from .roi_align import RoIAlign, roi_align
+
+__all__ = ['roi_align', 'RoIAlign']
diff --git a/lib/ops/roi_align/__pycache__/__init__.cpython-37.pyc b/lib/ops/roi_align/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..0743ac5
Binary files /dev/null and b/lib/ops/roi_align/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/roi_align/__pycache__/roi_align.cpython-37.pyc b/lib/ops/roi_align/__pycache__/roi_align.cpython-37.pyc
new file mode 100644
index 0000000..a9d0170
Binary files /dev/null and b/lib/ops/roi_align/__pycache__/roi_align.cpython-37.pyc differ
diff --git a/lib/ops/roi_align/gradcheck.py b/lib/ops/roi_align/gradcheck.py
new file mode 100644
index 0000000..136456b
--- /dev/null
+++ b/lib/ops/roi_align/gradcheck.py
@@ -0,0 +1,30 @@
+import os.path as osp
+import sys
+
+import numpy as np
+import torch
+from torch.autograd import gradcheck
+
+sys.path.append(osp.abspath(osp.join(__file__, '../../')))
+from roi_align import RoIAlign  # noqa: E402, isort:skip
+
+feat_size = 15
+spatial_scale = 1.0 / 8
+img_size = feat_size / spatial_scale
+num_imgs = 2
+num_rois = 20
+
+batch_ind = np.random.randint(num_imgs, size=(num_rois, 1))
+rois = np.random.rand(num_rois, 4) * img_size * 0.5
+rois[:, 2:] += img_size * 0.5
+rois = np.hstack((batch_ind, rois))
+
+feat = torch.randn(
+    num_imgs, 16, feat_size, feat_size, requires_grad=True, device='cuda:0')
+rois = torch.from_numpy(rois).float().cuda()
+inputs = (feat, rois)
+print('Gradcheck for roi align...')
+test = gradcheck(RoIAlign(3, spatial_scale), inputs, atol=1e-3, eps=1e-3)
+print(test)
+test = gradcheck(RoIAlign(3, spatial_scale, 2), inputs, atol=1e-3, eps=1e-3)
+print(test)
diff --git a/lib/ops/roi_align/roi_align.py b/lib/ops/roi_align/roi_align.py
new file mode 100644
index 0000000..a4cf244
--- /dev/null
+++ b/lib/ops/roi_align/roi_align.py
@@ -0,0 +1,87 @@
+import torch.nn as nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.modules.utils import _pair
+
+from . import roi_align_cuda
+
+
+class RoIAlignFunction(Function):
+
+    @staticmethod
+    def forward(ctx, features, rois, out_size, spatial_scale, sample_num=0):
+        out_h, out_w = _pair(out_size)
+        assert isinstance(out_h, int) and isinstance(out_w, int)
+        ctx.spatial_scale = spatial_scale
+        ctx.sample_num = sample_num
+        ctx.save_for_backward(rois)
+        ctx.feature_size = features.size()
+
+        batch_size, num_channels, data_height, data_width = features.size()
+        num_rois = rois.size(0)
+
+        output = features.new_zeros(num_rois, num_channels, out_h, out_w)
+        if features.is_cuda:
+            roi_align_cuda.forward(features, rois, out_h, out_w, spatial_scale,
+                                   sample_num, output)
+        else:
+            raise NotImplementedError
+
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        feature_size = ctx.feature_size
+        spatial_scale = ctx.spatial_scale
+        sample_num = ctx.sample_num
+        rois = ctx.saved_tensors[0]
+        assert (feature_size is not None and grad_output.is_cuda)
+
+        batch_size, num_channels, data_height, data_width = feature_size
+        out_w = grad_output.size(3)
+        out_h = grad_output.size(2)
+
+        grad_input = grad_rois = None
+        if ctx.needs_input_grad[0]:
+            grad_input = rois.new_zeros(batch_size, num_channels, data_height,
+                                        data_width)
+            roi_align_cuda.backward(grad_output.contiguous(), rois, out_h,
+                                    out_w, spatial_scale, sample_num,
+                                    grad_input)
+
+        return grad_input, grad_rois, None, None, None
+
+
+roi_align = RoIAlignFunction.apply
+
+
+class RoIAlign(nn.Module):
+
+    def __init__(self,
+                 out_size,
+                 spatial_scale,
+                 sample_num=0,
+                 use_torchvision=False):
+        super(RoIAlign, self).__init__()
+
+        self.out_size = _pair(out_size)
+        self.spatial_scale = float(spatial_scale)
+        self.sample_num = int(sample_num)
+        self.use_torchvision = use_torchvision
+
+    def forward(self, features, rois):
+        if self.use_torchvision:
+            from torchvision.ops import roi_align as tv_roi_align
+            return tv_roi_align(features, rois, self.out_size,
+                                self.spatial_scale, self.sample_num)
+        else:
+            return roi_align(features, rois, self.out_size, self.spatial_scale,
+                             self.sample_num)
+
+    def __repr__(self):
+        format_str = self.__class__.__name__
+        format_str += '(out_size={}, spatial_scale={}, sample_num={}'.format(
+            self.out_size, self.spatial_scale, self.sample_num)
+        format_str += ', use_torchvision={})'.format(self.use_torchvision)
+        return format_str
diff --git a/lib/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..19727f4
Binary files /dev/null and b/lib/ops/roi_align/roi_align_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/roi_align/src/roi_align_cuda.cpp b/lib/ops/roi_align/src/roi_align_cuda.cpp
new file mode 100644
index 0000000..06a73aa
--- /dev/null
+++ b/lib/ops/roi_align/src/roi_align_cuda.cpp
@@ -0,0 +1,85 @@
+#include <torch/extension.h>
+
+#include <cmath>
+#include <vector>
+
+int ROIAlignForwardLaucher(const at::Tensor features, const at::Tensor rois,
+                           const float spatial_scale, const int sample_num,
+                           const int channels, const int height,
+                           const int width, const int num_rois,
+                           const int pooled_height, const int pooled_width,
+                           at::Tensor output);
+
+int ROIAlignBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
+                            const float spatial_scale, const int sample_num,
+                            const int channels, const int height,
+                            const int width, const int num_rois,
+                            const int pooled_height, const int pooled_width,
+                            at::Tensor bottom_grad);
+
+#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
+#define CHECK_CONTIGUOUS(x) \
+  AT_CHECK(x.is_contiguous(), #x, " must be contiguous ")
+#define CHECK_INPUT(x) \
+  CHECK_CUDA(x);       \
+  CHECK_CONTIGUOUS(x)
+
+int roi_align_forward_cuda(at::Tensor features, at::Tensor rois,
+                           int pooled_height, int pooled_width,
+                           float spatial_scale, int sample_num,
+                           at::Tensor output) {
+  CHECK_INPUT(features);
+  CHECK_INPUT(rois);
+  CHECK_INPUT(output);
+
+  // Number of ROIs
+  int num_rois = rois.size(0);
+  int size_rois = rois.size(1);
+
+  if (size_rois != 5) {
+    printf("wrong roi size\n");
+    return 0;
+  }
+
+  int num_channels = features.size(1);
+  int data_height = features.size(2);
+  int data_width = features.size(3);
+
+  ROIAlignForwardLaucher(features, rois, spatial_scale, sample_num,
+                         num_channels, data_height, data_width, num_rois,
+                         pooled_height, pooled_width, output);
+
+  return 1;
+}
+
+int roi_align_backward_cuda(at::Tensor top_grad, at::Tensor rois,
+                            int pooled_height, int pooled_width,
+                            float spatial_scale, int sample_num,
+                            at::Tensor bottom_grad) {
+  CHECK_INPUT(top_grad);
+  CHECK_INPUT(rois);
+  CHECK_INPUT(bottom_grad);
+
+  // Number of ROIs
+  int num_rois = rois.size(0);
+  int size_rois = rois.size(1);
+  if (size_rois != 5) {
+    printf("wrong roi size\n");
+    return 0;
+  }
+
+  int num_channels = bottom_grad.size(1);
+  int data_height = bottom_grad.size(2);
+  int data_width = bottom_grad.size(3);
+
+  ROIAlignBackwardLaucher(top_grad, rois, spatial_scale, sample_num,
+                          num_channels, data_height, data_width, num_rois,
+                          pooled_height, pooled_width, bottom_grad);
+
+  return 1;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward", &roi_align_forward_cuda, "Roi_Align forward (CUDA)");
+  m.def("backward", &roi_align_backward_cuda, "Roi_Align backward (CUDA)");
+}
diff --git a/lib/ops/roi_align/src/roi_align_kernel.cu b/lib/ops/roi_align/src/roi_align_kernel.cu
new file mode 100644
index 0000000..6d3b279
--- /dev/null
+++ b/lib/ops/roi_align/src/roi_align_kernel.cu
@@ -0,0 +1,294 @@
+#include <ATen/ATen.h>
+#include <THC/THCAtomics.cuh>
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                            \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
+       i += blockDim.x * gridDim.x)
+
+#define THREADS_PER_BLOCK 1024
+
+inline int GET_BLOCKS(const int N) {
+  int optimal_block_num = (N + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+  int max_block_num = 65000;
+  return min(optimal_block_num, max_block_num);
+}
+
+template <typename scalar_t>
+__device__ scalar_t bilinear_interpolate(const scalar_t *bottom_data,
+                                         const int height, const int width,
+                                         scalar_t y, scalar_t x) {
+  // deal with cases that inverse elements are out of feature map boundary
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    return 0;
+  }
+
+  if (y <= 0) y = 0;
+  if (x <= 0) x = 0;
+
+  int y_low = (int)y;
+  int x_low = (int)x;
+  int y_high;
+  int x_high;
+
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = (scalar_t)y_low;
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = (scalar_t)x_low;
+  } else {
+    x_high = x_low + 1;
+  }
+
+  scalar_t ly = y - y_low;
+  scalar_t lx = x - x_low;
+  scalar_t hy = 1. - ly;
+  scalar_t hx = 1. - lx;
+  // do bilinear interpolation
+  scalar_t lt = bottom_data[y_low * width + x_low];
+  scalar_t rt = bottom_data[y_low * width + x_high];
+  scalar_t lb = bottom_data[y_high * width + x_low];
+  scalar_t rb = bottom_data[y_high * width + x_high];
+  scalar_t w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  scalar_t val = (w1 * lt + w2 * rt + w3 * lb + w4 * rb);
+
+  return val;
+}
+
+template <typename scalar_t>
+__global__ void ROIAlignForward(const int nthreads, const scalar_t *bottom_data,
+                                const scalar_t *bottom_rois,
+                                const scalar_t spatial_scale,
+                                const int sample_num, const int channels,
+                                const int height, const int width,
+                                const int pooled_height, const int pooled_width,
+                                scalar_t *top_data) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    // (n, c, ph, pw) is an element in the aligned output
+    int pw = index % pooled_width;
+    int ph = (index / pooled_width) % pooled_height;
+    int c = (index / pooled_width / pooled_height) % channels;
+    int n = index / pooled_width / pooled_height / channels;
+
+    const scalar_t *offset_bottom_rois = bottom_rois + n * 5;
+    int roi_batch_ind = offset_bottom_rois[0];
+    scalar_t roi_start_w = offset_bottom_rois[1] * spatial_scale;
+    scalar_t roi_start_h = offset_bottom_rois[2] * spatial_scale;
+    scalar_t roi_end_w = (offset_bottom_rois[3] + 1) * spatial_scale;
+    scalar_t roi_end_h = (offset_bottom_rois[4] + 1) * spatial_scale;
+
+    // Force malformed ROIs to be 1x1
+    scalar_t roi_width = fmaxf((scalar_t)roi_end_w - roi_start_w, 0.);
+    scalar_t roi_height = fmaxf((scalar_t)roi_end_h - roi_start_h, 0.);
+
+    scalar_t bin_size_h = roi_height / pooled_height;
+    scalar_t bin_size_w = roi_width / pooled_width;
+
+    const scalar_t *offset_bottom_data =
+        bottom_data + (roi_batch_ind * channels + c) * height * width;
+
+    int sample_num_h = (sample_num > 0)
+                           ? sample_num
+                           : ceil(roi_height / pooled_height);  // e.g., = 2
+    int sample_num_w =
+        (sample_num > 0) ? sample_num : ceil(roi_width / pooled_width);
+
+    scalar_t h = (scalar_t)(ph + 0.5) * bin_size_h + roi_start_h;
+    scalar_t w = (scalar_t)(pw + 0.5) * bin_size_w + roi_start_w;
+
+    int hstart = fminf(floor(h), height - 2);
+    int wstart = fminf(floor(w), width - 2);
+
+    scalar_t output_val = 0;
+    for (int iy = 0; iy < sample_num_h; iy++) {
+      const scalar_t y = roi_start_h + ph * bin_size_h +
+                         (scalar_t)(iy + scalar_t(.5f)) * bin_size_h /
+                             (scalar_t)(sample_num_h);
+      for (int ix = 0; ix < sample_num_w; ix++) {
+        const scalar_t x = roi_start_w + pw * bin_size_w +
+                           (scalar_t)(ix + scalar_t(.5f)) * bin_size_w /
+                               (scalar_t)(sample_num_w);
+        scalar_t val = bilinear_interpolate<scalar_t>(offset_bottom_data,
+                                                      height, width, y, x);
+        output_val += val;
+      }
+    }
+    output_val /= (sample_num_h * sample_num_w);
+    top_data[index] = output_val;
+  }
+}
+
+int ROIAlignForwardLaucher(const at::Tensor features, const at::Tensor rois,
+                           const float spatial_scale, const int sample_num,
+                           const int channels, const int height,
+                           const int width, const int num_rois,
+                           const int pooled_height, const int pooled_width,
+                           at::Tensor output) {
+  const int output_size = num_rois * pooled_height * pooled_width * channels;
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      features.scalar_type(), "ROIAlignLaucherForward", ([&] {
+        const scalar_t *bottom_data = features.data<scalar_t>();
+        const scalar_t *rois_data = rois.data<scalar_t>();
+        scalar_t *top_data = output.data<scalar_t>();
+
+        ROIAlignForward<scalar_t>
+            <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK>>>(
+                output_size, bottom_data, rois_data, scalar_t(spatial_scale),
+                sample_num, channels, height, width, pooled_height,
+                pooled_width, top_data);
+      }));
+  THCudaCheck(cudaGetLastError());
+  return 1;
+}
+
+template <typename scalar_t>
+__device__ void bilinear_interpolate_gradient(const int height, const int width,
+                                              scalar_t y, scalar_t x,
+                                              scalar_t &w1, scalar_t &w2,
+                                              scalar_t &w3, scalar_t &w4,
+                                              int &x_low, int &x_high,
+                                              int &y_low, int &y_high) {
+  // deal with cases that inverse elements are out of feature map boundary
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    w1 = w2 = w3 = w4 = 0.;
+    x_low = x_high = y_low = y_high = -1;
+    return;
+  }
+
+  if (y <= 0) y = 0;
+  if (x <= 0) x = 0;
+
+  y_low = (int)y;
+  x_low = (int)x;
+
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = (scalar_t)y_low;
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = (scalar_t)x_low;
+  } else {
+    x_high = x_low + 1;
+  }
+
+  scalar_t ly = y - y_low;
+  scalar_t lx = x - x_low;
+  scalar_t hy = 1. - ly;
+  scalar_t hx = 1. - lx;
+
+  w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  return;
+}
+
+template <typename scalar_t>
+__global__ void ROIAlignBackward(
+    const int nthreads, const scalar_t *top_diff, const scalar_t *bottom_rois,
+    const scalar_t spatial_scale, const int sample_num, const int channels,
+    const int height, const int width, const int pooled_height,
+    const int pooled_width, scalar_t *bottom_diff) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    // (n, c, ph, pw) is an element in the aligned output
+    int pw = index % pooled_width;
+    int ph = (index / pooled_width) % pooled_height;
+    int c = (index / pooled_width / pooled_height) % channels;
+    int n = index / pooled_width / pooled_height / channels;
+
+    const scalar_t *offset_bottom_rois = bottom_rois + n * 5;
+    int roi_batch_ind = offset_bottom_rois[0];
+    scalar_t roi_start_w = offset_bottom_rois[1] * spatial_scale;
+    scalar_t roi_start_h = offset_bottom_rois[2] * spatial_scale;
+    scalar_t roi_end_w = (offset_bottom_rois[3] + 1) * spatial_scale;
+    scalar_t roi_end_h = (offset_bottom_rois[4] + 1) * spatial_scale;
+
+    // Force malformed ROIs to be 1x1
+    scalar_t roi_width = fmaxf((scalar_t)roi_end_w - roi_start_w, 0.);
+    scalar_t roi_height = fmaxf((scalar_t)roi_end_h - roi_start_h, 0.);
+
+    scalar_t bin_size_h = roi_height / pooled_height;
+    scalar_t bin_size_w = roi_width / pooled_width;
+
+    scalar_t *offset_bottom_diff =
+        bottom_diff + (roi_batch_ind * channels + c) * height * width;
+    int offset_top = (n * channels + c) * pooled_height * pooled_width +
+                     ph * pooled_width + pw;
+    scalar_t offset_top_diff = top_diff[offset_top];
+
+    int sample_num_h = (sample_num > 0)
+                           ? sample_num
+                           : ceil(roi_height / pooled_height);  // e.g., = 2
+    int sample_num_w =
+        (sample_num > 0) ? sample_num : ceil(roi_width / pooled_width);
+
+    const scalar_t count = (scalar_t)(sample_num_h * sample_num_w);
+
+    scalar_t h = (scalar_t)(ph + 0.5) * bin_size_h + roi_start_h;
+    scalar_t w = (scalar_t)(pw + 0.5) * bin_size_w + roi_start_w;
+
+    int hstart = fminf(floor(h), height - 2);
+    int wstart = fminf(floor(w), width - 2);
+
+    for (int iy = 0; iy < sample_num_h; iy++) {
+      const scalar_t y =
+          roi_start_h + ph * bin_size_h +
+          (scalar_t)(iy + .5f) * bin_size_h / (scalar_t)(sample_num_h);
+      for (int ix = 0; ix < sample_num_w; ix++) {
+        const scalar_t x =
+            roi_start_w + pw * bin_size_w +
+            (scalar_t)(ix + .5f) * bin_size_w / (scalar_t)(sample_num_w);
+        scalar_t w1, w2, w3, w4;
+        int x_low, x_high, y_low, y_high;
+
+        bilinear_interpolate_gradient<scalar_t>(
+            height, width, y, x, w1, w2, w3, w4, x_low, x_high, y_low, y_high);
+        scalar_t g1 = offset_top_diff * w1 / count;
+        scalar_t g2 = offset_top_diff * w2 / count;
+        scalar_t g3 = offset_top_diff * w3 / count;
+        scalar_t g4 = offset_top_diff * w4 / count;
+        if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
+          atomicAdd(offset_bottom_diff + y_low * width + x_low, g1);
+          atomicAdd(offset_bottom_diff + y_low * width + x_high, g2);
+          atomicAdd(offset_bottom_diff + y_high * width + x_low, g3);
+          atomicAdd(offset_bottom_diff + y_high * width + x_high, g4);
+        }
+      }
+    }
+  }
+}
+
+int ROIAlignBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
+                            const float spatial_scale, const int sample_num,
+                            const int channels, const int height,
+                            const int width, const int num_rois,
+                            const int pooled_height, const int pooled_width,
+                            at::Tensor bottom_grad) {
+  const int output_size = num_rois * pooled_height * pooled_width * channels;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      top_grad.scalar_type(), "ROIAlignLaucherBackward", ([&] {
+        const scalar_t *top_diff = top_grad.data<scalar_t>();
+        const scalar_t *rois_data = rois.data<scalar_t>();
+        scalar_t *bottom_diff = bottom_grad.data<scalar_t>();
+        if (sizeof(scalar_t) == sizeof(double)) {
+          fprintf(stderr, "double is not supported\n");
+          exit(-1);
+        }
+
+        ROIAlignBackward<scalar_t>
+            <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK>>>(
+                output_size, top_diff, rois_data, spatial_scale, sample_num,
+                channels, height, width, pooled_height, pooled_width,
+                bottom_diff);
+      }));
+  THCudaCheck(cudaGetLastError());
+  return 1;
+}
diff --git a/lib/ops/roi_crop/__init__.py b/lib/ops/roi_crop/__init__.py
new file mode 100644
index 0000000..e1410ed
--- /dev/null
+++ b/lib/ops/roi_crop/__init__.py
@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Sep 27 14:18:45 2019
+
+@author: vasgaoweithu
+"""
+
+from .roi_crop import RoICrop
+
+__all__ = ['RoICrop']
\ No newline at end of file
diff --git a/lib/ops/roi_crop/__pycache__/__init__.cpython-37.pyc b/lib/ops/roi_crop/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..7d08353
Binary files /dev/null and b/lib/ops/roi_crop/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/roi_crop/__pycache__/roi_crop.cpython-37.pyc b/lib/ops/roi_crop/__pycache__/roi_crop.cpython-37.pyc
new file mode 100644
index 0000000..7c340cb
Binary files /dev/null and b/lib/ops/roi_crop/__pycache__/roi_crop.cpython-37.pyc differ
diff --git a/lib/ops/roi_crop/roi_crop.py b/lib/ops/roi_crop/roi_crop.py
new file mode 100644
index 0000000..666c5b7
--- /dev/null
+++ b/lib/ops/roi_crop/roi_crop.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Sep 27 13:56:53 2019
+
+@author: vasgaoweithu
+"""
+
+import torch
+import torch.nn as nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.modules.utils import _pair
+
+from . import roi_crop_cpu
+from . import roi_crop_cuda
+
+class RoICropFunction(Function):
+    @staticmethod
+    def forward(ctx, input1, input2):
+        ctx.input1 = input1.clone()
+        ctx.input2 = input2.clone()
+        output = input2.new(input2.size()[0], input1.size()[1], input2.size()[1], input2.size()[2]).zero_()
+        assert output.get_device() == input1.get_device(), "output and input1 must on the same device"
+        assert output.get_device() == input2.get_device(), "output and input2 must on the same device"
+        if input1.is_cuda:
+            roi_crop_cuda.forward(input1, input2, output)
+        else:
+            roi_crop_cpu.forward(input1, input2, output)
+        
+        return output
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        grad_input1 = ctx.input1.new(ctx.input1.size()).zero_()
+        grad_input2 = ctx.input2.new(ctx.input2.size()).zero_()
+        if grad_output.is_cuda:
+            roi_crop_cuda.backward(ctx.input1, ctx.input2, grad_input1, grad_input2, grad_output)
+        else:
+            roi_crop_cpu.backward(ctx.input1, ctx.input2, grad_input1, grad_input2, grad_output)
+        
+        return grad_input1, grad_input2
+    
+roi_crop = RoICropFunction.apply
+
+class RoICrop(nn.Module):
+    def __init__(self, layout='BHWD'):
+        super(RoICrop, self).__init__()
+    def forward(self, input1, input2):
+        return roi_crop(input1, input2)
\ No newline at end of file
diff --git a/lib/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so b/lib/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..8fb3628
Binary files /dev/null and b/lib/ops/roi_crop/roi_crop_cpu.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..d2ea63a
Binary files /dev/null and b/lib/ops/roi_crop/roi_crop_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/roi_crop/src/roi_crop_cpu.cpp b/lib/ops/roi_crop/src/roi_crop_cpu.cpp
new file mode 100644
index 0000000..32a223d
--- /dev/null
+++ b/lib/ops/roi_crop/src/roi_crop_cpu.cpp
@@ -0,0 +1,489 @@
+#include <TH/TH.h>
+#include<torch/extension.h>
+#include <stdbool.h>
+#include <stdio.h>
+
+#define real float
+
+int BilinearSamplerBHWD_updateOutput(THFloatTensor *inputImages, THFloatTensor *grids, THFloatTensor *output)
+{
+
+  int batchsize = THFloatTensor_size(inputImages, 0);
+  int inputImages_height = THFloatTensor_size(inputImages, 1);
+  int inputImages_width = THFloatTensor_size(inputImages, 2);
+  int output_height = THFloatTensor_size(output, 1);
+  int output_width = THFloatTensor_size(output, 2);
+  int inputImages_channels = THFloatTensor_size(inputImages, 3);
+
+  int output_strideBatch = THFloatTensor_stride(output, 0);
+  int output_strideHeight = THFloatTensor_stride(output, 1);
+  int output_strideWidth = THFloatTensor_stride(output, 2);
+
+  int inputImages_strideBatch = THFloatTensor_stride(inputImages, 0);
+  int inputImages_strideHeight = THFloatTensor_stride(inputImages, 1);
+  int inputImages_strideWidth = THFloatTensor_stride(inputImages, 2);
+
+  int grids_strideBatch = THFloatTensor_stride(grids, 0);
+  int grids_strideHeight = THFloatTensor_stride(grids, 1);
+  int grids_strideWidth = THFloatTensor_stride(grids, 2);
+
+  real *inputImages_data, *output_data, *grids_data;
+  inputImages_data = THFloatTensor_data(inputImages);
+  output_data = THFloatTensor_data(output);
+  grids_data = THFloatTensor_data(grids);
+
+  int b, yOut, xOut;
+
+  for(b=0; b < batchsize; b++)
+  {
+    for(yOut=0; yOut < output_height; yOut++)
+    {
+      for(xOut=0; xOut < output_width; xOut++)
+      {
+        //read the grid
+        real yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+        real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1];
+
+        // get the weights for interpolation
+        int yInTopLeft, xInTopLeft;
+        real yWeightTopLeft, xWeightTopLeft;
+
+        real xcoord = (xf + 1) * (inputImages_width - 1) / 2;
+        xInTopLeft = floor(xcoord);
+        xWeightTopLeft = 1 - (xcoord - xInTopLeft);
+
+        real ycoord = (yf + 1) * (inputImages_height - 1) / 2;
+        yInTopLeft = floor(ycoord);
+        yWeightTopLeft = 1 - (ycoord - yInTopLeft);
+
+
+
+        const int outAddress = output_strideBatch * b + output_strideHeight * yOut + output_strideWidth * xOut;
+        const int inTopLeftAddress = inputImages_strideBatch * b + inputImages_strideHeight * yInTopLeft + inputImages_strideWidth * xInTopLeft;
+        const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+        const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+        const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+        real v=0;
+        real inTopLeft=0;
+        real inTopRight=0;
+        real inBottomLeft=0;
+        real inBottomRight=0;
+
+        // we are careful with the boundaries
+        bool topLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool topRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool bottomLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+        bool bottomRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+
+        int t;
+        // interpolation happens here
+        for(t=0; t<inputImages_channels; t++)
+        {
+           if(topLeftIsIn) inTopLeft = inputImages_data[inTopLeftAddress + t];
+           if(topRightIsIn) inTopRight = inputImages_data[inTopRightAddress + t];
+           if(bottomLeftIsIn) inBottomLeft = inputImages_data[inBottomLeftAddress + t];
+           if(bottomRightIsIn) inBottomRight = inputImages_data[inBottomRightAddress + t];
+
+           v = xWeightTopLeft * yWeightTopLeft * inTopLeft
+             + (1 - xWeightTopLeft) * yWeightTopLeft * inTopRight
+             + xWeightTopLeft * (1 - yWeightTopLeft) * inBottomLeft
+             + (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * inBottomRight;
+
+           output_data[outAddress + t] = v;
+        }
+
+      }
+    }
+  }
+
+  return 1;
+}
+
+
+
+int BilinearSamplerBHWD_updateGradInput(THFloatTensor *inputImages, THFloatTensor *grids, THFloatTensor *gradInputImages,
+                                        THFloatTensor *gradGrids, THFloatTensor *gradOutput)
+{
+  bool onlyGrid=false;
+
+  int batchsize = THFloatTensor_size(inputImages, 0);
+  int inputImages_height = THFloatTensor_size(inputImages, 1);
+  int inputImages_width = THFloatTensor_size(inputImages, 2);
+  int gradOutput_height = THFloatTensor_size(gradOutput, 1);
+  int gradOutput_width = THFloatTensor_size(gradOutput, 2);
+  int inputImages_channels = THFloatTensor_size(inputImages, 3);
+
+  int gradOutput_strideBatch = THFloatTensor_stride(gradOutput, 0);
+  int gradOutput_strideHeight = THFloatTensor_stride(gradOutput, 1);
+  int gradOutput_strideWidth = THFloatTensor_stride(gradOutput, 2);
+
+  int inputImages_strideBatch = THFloatTensor_stride(inputImages, 0);
+  int inputImages_strideHeight = THFloatTensor_stride(inputImages, 1);
+  int inputImages_strideWidth = THFloatTensor_stride(inputImages, 2);
+
+  int gradInputImages_strideBatch = THFloatTensor_stride(gradInputImages, 0);
+  int gradInputImages_strideHeight = THFloatTensor_stride(gradInputImages, 1);
+  int gradInputImages_strideWidth = THFloatTensor_stride(gradInputImages, 2);
+
+  int grids_strideBatch = THFloatTensor_stride(grids, 0);
+  int grids_strideHeight = THFloatTensor_stride(grids, 1);
+  int grids_strideWidth = THFloatTensor_stride(grids, 2);
+
+  int gradGrids_strideBatch = THFloatTensor_stride(gradGrids, 0);
+  int gradGrids_strideHeight = THFloatTensor_stride(gradGrids, 1);
+  int gradGrids_strideWidth = THFloatTensor_stride(gradGrids, 2);
+
+  real *inputImages_data, *gradOutput_data, *grids_data, *gradGrids_data, *gradInputImages_data;
+  inputImages_data = THFloatTensor_data(inputImages);
+  gradOutput_data = THFloatTensor_data(gradOutput);
+  grids_data = THFloatTensor_data(grids);
+  gradGrids_data = THFloatTensor_data(gradGrids);
+  gradInputImages_data = THFloatTensor_data(gradInputImages);
+
+  int b, yOut, xOut;
+
+  for(b=0; b < batchsize; b++)
+  {
+    for(yOut=0; yOut < gradOutput_height; yOut++)
+    {
+      for(xOut=0; xOut < gradOutput_width; xOut++)
+      {
+        //read the grid
+        real yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+        real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1];
+
+        // get the weights for interpolation
+        int yInTopLeft, xInTopLeft;
+        real yWeightTopLeft, xWeightTopLeft;
+
+        real xcoord = (xf + 1) * (inputImages_width - 1) / 2;
+        xInTopLeft = floor(xcoord);
+        xWeightTopLeft = 1 - (xcoord - xInTopLeft);
+
+        real ycoord = (yf + 1) * (inputImages_height - 1) / 2;
+        yInTopLeft = floor(ycoord);
+        yWeightTopLeft = 1 - (ycoord - yInTopLeft);
+
+
+        const int inTopLeftAddress = inputImages_strideBatch * b + inputImages_strideHeight * yInTopLeft + inputImages_strideWidth * xInTopLeft;
+        const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+        const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+        const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+        const int gradInputImagesTopLeftAddress = gradInputImages_strideBatch * b + gradInputImages_strideHeight * yInTopLeft + gradInputImages_strideWidth * xInTopLeft;
+        const int gradInputImagesTopRightAddress = gradInputImagesTopLeftAddress + gradInputImages_strideWidth;
+        const int gradInputImagesBottomLeftAddress = gradInputImagesTopLeftAddress + gradInputImages_strideHeight;
+        const int gradInputImagesBottomRightAddress = gradInputImagesBottomLeftAddress + gradInputImages_strideWidth;
+
+        const int gradOutputAddress = gradOutput_strideBatch * b + gradOutput_strideHeight * yOut + gradOutput_strideWidth * xOut;
+
+        real topLeftDotProduct = 0;
+        real topRightDotProduct = 0;
+        real bottomLeftDotProduct = 0;
+        real bottomRightDotProduct = 0;
+
+        // we are careful with the boundaries
+        bool topLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool topRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool bottomLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+        bool bottomRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+
+        int t;
+
+        for(t=0; t<inputImages_channels; t++)
+        {
+           real gradOutValue = gradOutput_data[gradOutputAddress + t];
+           if(topLeftIsIn)
+           {
+              real inTopLeft = inputImages_data[inTopLeftAddress + t];
+              topLeftDotProduct += inTopLeft * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesTopLeftAddress + t] += xWeightTopLeft * yWeightTopLeft * gradOutValue;
+           }
+
+           if(topRightIsIn)
+           {
+              real inTopRight = inputImages_data[inTopRightAddress + t];
+              topRightDotProduct += inTopRight * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesTopRightAddress + t] += (1 - xWeightTopLeft) * yWeightTopLeft * gradOutValue;
+           }
+
+           if(bottomLeftIsIn)
+           {
+              real inBottomLeft = inputImages_data[inBottomLeftAddress + t];
+              bottomLeftDotProduct += inBottomLeft * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesBottomLeftAddress + t] += xWeightTopLeft * (1 - yWeightTopLeft) * gradOutValue;
+           }
+
+           if(bottomRightIsIn)
+           {
+              real inBottomRight = inputImages_data[inBottomRightAddress + t];
+              bottomRightDotProduct += inBottomRight * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesBottomRightAddress + t] += (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * gradOutValue;
+           }
+        }
+
+        yf = - xWeightTopLeft * topLeftDotProduct + xWeightTopLeft * bottomLeftDotProduct - (1-xWeightTopLeft) * topRightDotProduct + (1-xWeightTopLeft) * bottomRightDotProduct;
+        xf = - yWeightTopLeft * topLeftDotProduct + yWeightTopLeft * topRightDotProduct - (1-yWeightTopLeft) * bottomLeftDotProduct + (1-yWeightTopLeft) * bottomRightDotProduct;
+
+        gradGrids_data[b*gradGrids_strideBatch + yOut*gradGrids_strideHeight + xOut*gradGrids_strideWidth] = yf * (inputImages_height-1) / 2;
+        gradGrids_data[b*gradGrids_strideBatch + yOut*gradGrids_strideHeight + xOut*gradGrids_strideWidth + 1] = xf * (inputImages_width-1) / 2;
+
+      }
+    }
+  }
+
+  return 1;
+}
+
+
+int BilinearSamplerBCHW_updateOutput(THFloatTensor *inputImages, THFloatTensor *grids, THFloatTensor *output)
+{
+
+  int batchsize = THFloatTensor_size(inputImages, 0);
+  int inputImages_height = THFloatTensor_size(inputImages, 2);
+  int inputImages_width = THFloatTensor_size(inputImages, 3);
+  
+  int output_height = THFloatTensor_size(output, 2);
+  int output_width = THFloatTensor_size(output, 3);
+  int inputImages_channels = THFloatTensor_size(inputImages, 1);
+
+  int output_strideBatch = THFloatTensor_stride(output, 0);
+  int output_strideHeight = THFloatTensor_stride(output, 2);
+  int output_strideWidth = THFloatTensor_stride(output, 3);  
+  int output_strideChannel = THFloatTensor_stride(output, 1);
+    
+  int inputImages_strideBatch = THFloatTensor_stride(inputImages, 0);
+  int inputImages_strideHeight = THFloatTensor_stride(inputImages, 2);
+  int inputImages_strideWidth = THFloatTensor_stride(inputImages, 3);
+  int inputImages_strideChannel = THFloatTensor_stride(inputImages, 1);
+
+  int grids_strideBatch = THFloatTensor_stride(grids, 0);
+  int grids_strideHeight = THFloatTensor_stride(grids, 2);
+  int grids_strideWidth = THFloatTensor_stride(grids, 3);
+  int grids_strideChannel = THFloatTensor_stride(grids, 1);
+
+
+  real *inputImages_data, *output_data, *grids_data;
+  inputImages_data = THFloatTensor_data(inputImages);
+  output_data = THFloatTensor_data(output);
+  grids_data = THFloatTensor_data(grids);
+
+  int b, yOut, xOut;
+
+  for(b=0; b < batchsize; b++)
+  {
+    for(yOut=0; yOut < output_height; yOut++)
+    {
+      for(xOut=0; xOut < output_width; xOut++)
+      {
+        //read the grid
+        
+        real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + grids_strideChannel];
+        real yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+
+        // get the weights for interpolation
+        int yInTopLeft, xInTopLeft;
+        real yWeightTopLeft, xWeightTopLeft;
+
+        real xcoord = (xf + 1) * (inputImages_width - 1) / 2;
+        xInTopLeft = floor(xcoord);
+        xWeightTopLeft = 1 - (xcoord - xInTopLeft);
+
+        real ycoord = (yf + 1) * (inputImages_height - 1) / 2;
+        yInTopLeft = floor(ycoord);
+        yWeightTopLeft = 1 - (ycoord - yInTopLeft);
+
+
+
+        const int outAddress = output_strideBatch * b + output_strideHeight * yOut + output_strideWidth * xOut;
+        const int inTopLeftAddress = inputImages_strideBatch * b + inputImages_strideHeight * yInTopLeft + inputImages_strideWidth * xInTopLeft;
+        const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+        const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+        const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+        real v=0;
+        real inTopLeft=0;
+        real inTopRight=0;
+        real inBottomLeft=0;
+        real inBottomRight=0;
+
+        // we are careful with the boundaries
+        bool topLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool topRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool bottomLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+        bool bottomRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+
+        int t;
+        // interpolation happens here
+        for(t=0; t<inputImages_channels; t++)
+        {
+           if(topLeftIsIn) inTopLeft = inputImages_data[inTopLeftAddress + t * inputImages_strideChannel];
+           if(topRightIsIn) inTopRight = inputImages_data[inTopRightAddress + t * inputImages_strideChannel];
+           if(bottomLeftIsIn) inBottomLeft = inputImages_data[inBottomLeftAddress + t * inputImages_strideChannel];
+           if(bottomRightIsIn) inBottomRight = inputImages_data[inBottomRightAddress + t * inputImages_strideChannel];
+
+           v = xWeightTopLeft * yWeightTopLeft * inTopLeft
+             + (1 - xWeightTopLeft) * yWeightTopLeft * inTopRight
+             + xWeightTopLeft * (1 - yWeightTopLeft) * inBottomLeft
+             + (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * inBottomRight;
+
+           output_data[outAddress + t * output_strideChannel] = v;
+        }
+
+      }
+    }
+  }
+
+  return 1;
+}
+
+
+
+int BilinearSamplerBCHW_updateGradInput(THFloatTensor *inputImages, THFloatTensor *grids, THFloatTensor *gradInputImages,
+                                        THFloatTensor *gradGrids, THFloatTensor *gradOutput)
+{
+  bool onlyGrid=false;
+
+  int batchsize = THFloatTensor_size(inputImages, 0);
+  int inputImages_height = THFloatTensor_size(inputImages, 2);
+  int inputImages_width = THFloatTensor_size(inputImages, 3);
+  int gradOutput_height = THFloatTensor_size(gradOutput, 2);
+  int gradOutput_width = THFloatTensor_size(gradOutput, 3);
+  int inputImages_channels = THFloatTensor_size(inputImages, 1);
+
+  int gradOutput_strideBatch = THFloatTensor_stride(gradOutput, 0);
+  int gradOutput_strideHeight = THFloatTensor_stride(gradOutput, 2);
+  int gradOutput_strideWidth = THFloatTensor_stride(gradOutput, 3);
+  int gradOutput_strideChannel = THFloatTensor_stride(gradOutput, 1);
+
+  int inputImages_strideBatch = THFloatTensor_stride(inputImages, 0);
+  int inputImages_strideHeight = THFloatTensor_stride(inputImages, 2);
+  int inputImages_strideWidth = THFloatTensor_stride(inputImages, 3);
+  int inputImages_strideChannel = THFloatTensor_stride(inputImages, 1);
+    
+  int gradInputImages_strideBatch = THFloatTensor_stride(gradInputImages, 0);
+  int gradInputImages_strideHeight = THFloatTensor_stride(gradInputImages, 2);
+  int gradInputImages_strideWidth = THFloatTensor_stride(gradInputImages, 3);
+  int gradInputImages_strideChannel = THFloatTensor_stride(gradInputImages, 1);
+
+  int grids_strideBatch = THFloatTensor_stride(grids, 0);
+  int grids_strideHeight = THFloatTensor_stride(grids, 2);
+  int grids_strideWidth = THFloatTensor_stride(grids, 3);
+  int grids_strideChannel = THFloatTensor_stride(grids, 1);
+
+  int gradGrids_strideBatch = THFloatTensor_stride(gradGrids, 0);
+  int gradGrids_strideHeight = THFloatTensor_stride(gradGrids, 2);
+  int gradGrids_strideWidth = THFloatTensor_stride(gradGrids, 3);
+  int gradGrids_strideChannel = THFloatTensor_stride(gradGrids, 1);
+
+  real *inputImages_data, *gradOutput_data, *grids_data, *gradGrids_data, *gradInputImages_data;
+  inputImages_data = THFloatTensor_data(inputImages);
+  gradOutput_data = THFloatTensor_data(gradOutput);
+  grids_data = THFloatTensor_data(grids);
+  gradGrids_data = THFloatTensor_data(gradGrids);
+  gradInputImages_data = THFloatTensor_data(gradInputImages);
+
+  int b, yOut, xOut;
+
+  for(b=0; b < batchsize; b++)
+  {
+    for(yOut=0; yOut < gradOutput_height; yOut++)
+    {
+      for(xOut=0; xOut < gradOutput_width; xOut++)
+      {
+        //read the grid
+        real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + grids_strideChannel];
+        real yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+        
+        // get the weights for interpolation
+        int yInTopLeft, xInTopLeft;
+        real yWeightTopLeft, xWeightTopLeft;
+
+        real xcoord = (xf + 1) * (inputImages_width - 1) / 2;
+        xInTopLeft = floor(xcoord);
+        xWeightTopLeft = 1 - (xcoord - xInTopLeft);
+
+        real ycoord = (yf + 1) * (inputImages_height - 1) / 2;
+        yInTopLeft = floor(ycoord);
+        yWeightTopLeft = 1 - (ycoord - yInTopLeft);
+
+
+        const int inTopLeftAddress = inputImages_strideBatch * b + inputImages_strideHeight * yInTopLeft + inputImages_strideWidth * xInTopLeft;
+        const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+        const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+        const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+        const int gradInputImagesTopLeftAddress = gradInputImages_strideBatch * b + gradInputImages_strideHeight * yInTopLeft + gradInputImages_strideWidth * xInTopLeft;
+        const int gradInputImagesTopRightAddress = gradInputImagesTopLeftAddress + gradInputImages_strideWidth;
+        const int gradInputImagesBottomLeftAddress = gradInputImagesTopLeftAddress + gradInputImages_strideHeight;
+        const int gradInputImagesBottomRightAddress = gradInputImagesBottomLeftAddress + gradInputImages_strideWidth;
+
+        const int gradOutputAddress = gradOutput_strideBatch * b + gradOutput_strideHeight * yOut + gradOutput_strideWidth * xOut;
+
+        real topLeftDotProduct = 0;
+        real topRightDotProduct = 0;
+        real bottomLeftDotProduct = 0;
+        real bottomRightDotProduct = 0;
+
+        // we are careful with the boundaries
+        bool topLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool topRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft >= 0 && yInTopLeft <= inputImages_height-1;
+        bool bottomLeftIsIn = xInTopLeft >= 0 && xInTopLeft <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+        bool bottomRightIsIn = xInTopLeft+1 >= 0 && xInTopLeft+1 <= inputImages_width-1 && yInTopLeft+1 >= 0 && yInTopLeft+1 <= inputImages_height-1;
+
+        int t;
+
+        for(t=0; t<inputImages_channels; t++)
+        {
+           real gradOutValue = gradOutput_data[gradOutputAddress + t * gradOutput_strideChannel];
+           if(topLeftIsIn)
+           {
+              real inTopLeft = inputImages_data[inTopLeftAddress + t * inputImages_strideChannel];
+              topLeftDotProduct += inTopLeft * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesTopLeftAddress + t * gradInputImages_strideChannel] += xWeightTopLeft * yWeightTopLeft * gradOutValue;
+           }
+
+           if(topRightIsIn)
+           {
+              real inTopRight = inputImages_data[inTopRightAddress + t * inputImages_strideChannel];
+              topRightDotProduct += inTopRight * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesTopRightAddress + t * gradInputImages_strideChannel] += (1 - xWeightTopLeft) * yWeightTopLeft * gradOutValue;
+           }
+
+           if(bottomLeftIsIn)
+           {
+              real inBottomLeft = inputImages_data[inBottomLeftAddress + t * inputImages_strideChannel];
+              bottomLeftDotProduct += inBottomLeft * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesBottomLeftAddress + t * gradInputImages_strideChannel] += xWeightTopLeft * (1 - yWeightTopLeft) * gradOutValue;
+           }
+
+           if(bottomRightIsIn)
+           {
+              real inBottomRight = inputImages_data[inBottomRightAddress + t * inputImages_strideChannel];
+              bottomRightDotProduct += inBottomRight * gradOutValue;
+              if(!onlyGrid) gradInputImages_data[gradInputImagesBottomRightAddress + t * gradInputImages_strideChannel] += (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * gradOutValue;
+           }
+        }
+
+        xf = - yWeightTopLeft * topLeftDotProduct + yWeightTopLeft * topRightDotProduct - (1-yWeightTopLeft) * bottomLeftDotProduct + (1-yWeightTopLeft) * bottomRightDotProduct;
+          
+        yf = - xWeightTopLeft * topLeftDotProduct + xWeightTopLeft * bottomLeftDotProduct - (1-xWeightTopLeft) * topRightDotProduct + (1-xWeightTopLeft) * bottomRightDotProduct;
+        
+
+        gradGrids_data[b*gradGrids_strideBatch + yOut*gradGrids_strideHeight + xOut*gradGrids_strideWidth + gradGrids_strideChannel] = xf * (inputImages_width-1) / 2;
+          
+        gradGrids_data[b*gradGrids_strideBatch + yOut*gradGrids_strideHeight + xOut*gradGrids_strideWidth] = yf * (inputImages_height-1) / 2;
+        
+
+      }
+    }
+  }
+
+  return 1;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward", &BilinearSamplerBCHW_updateOutput, "roi crop forward (CPU)");
+  m.def("backward", &BilinearSamplerBCHW_updateGradInput, "roi crop backward (CPU)");
+}
diff --git a/lib/ops/roi_crop/src/roi_crop_cuda.cpp b/lib/ops/roi_crop/src/roi_crop_cuda.cpp
new file mode 100644
index 0000000..2f3ec5f
--- /dev/null
+++ b/lib/ops/roi_crop/src/roi_crop_cuda.cpp
@@ -0,0 +1,112 @@
+#include <THC/THC.h>
+#include<torch/extension.h>
+#include <stdbool.h>
+#include <stdio.h>
+
+#define real float
+
+int BilinearSamplerBHWD_updateOutput_cuda_kernel(int oc, int ow, int oh, int ob, int ic, int ih, int iw, int ib,
+                                                 float *inputImages, int isb, int isc, int ish, int isw,
+                                                 float *grids, int gsb, int gsc, int gsh, int gsw,
+                                                 float *output, int osb, int osc, int osh, int osw,
+                                                 cudaStream_t stream);
+
+int BilinearSamplerBHWD_updateGradInput_cuda_kernel(int goc, int gow, int goh, int gob, int ic, int ih, int iw, int ib, 
+                                                    float *inputImages, int isb, int isc, int ish, int isw,
+                                                    float *grids, int gsb, int gsc, int gsh, int gsw,
+                                                    float *gradInputImages, int gisb, int gisc, int gish, int gisw,
+                                                    float *gradGrids, int ggsb, int ggsc, int ggsh, int ggsw,
+                                                    float *gradOutput, int gosb, int gosc, int gosh, int gosw,
+                                                    cudaStream_t stream);
+
+// this symbol will be resolved automatically from PyTorch libs
+extern THCState *state;
+
+// Bilinear sampling is done in BHWD (coalescing is not obvious in BDHW)
+// we assume BHWD format in inputImages
+// we assume BHW(YX) format on grids
+
+int BilinearSamplerBHWD_updateOutput_cuda(THCudaTensor *inputImages, THCudaTensor *grids, THCudaTensor *output){
+  int success = 0;
+  success = BilinearSamplerBHWD_updateOutput_cuda_kernel(THCudaTensor_size(state, output, 1),
+							 THCudaTensor_size(state, output, 3),
+							 THCudaTensor_size(state, output, 2),
+							 THCudaTensor_size(state, output, 0),
+							 THCudaTensor_size(state, inputImages, 1),
+							 THCudaTensor_size(state, inputImages, 2),
+							 THCudaTensor_size(state, inputImages, 3),
+							 THCudaTensor_size(state, inputImages, 0),
+							 THCudaTensor_data(state, inputImages),
+							 THCudaTensor_stride(state, inputImages, 0),
+							 THCudaTensor_stride(state, inputImages, 1),
+							 THCudaTensor_stride(state, inputImages, 2),
+							 THCudaTensor_stride(state, inputImages, 3),
+							 THCudaTensor_data(state, grids),
+							 THCudaTensor_stride(state, grids, 0),
+							 THCudaTensor_stride(state, grids, 3),
+							 THCudaTensor_stride(state, grids, 1),
+							 THCudaTensor_stride(state, grids, 2),
+							 THCudaTensor_data(state, output),
+							 THCudaTensor_stride(state, output, 0),
+							 THCudaTensor_stride(state, output, 1),
+							 THCudaTensor_stride(state, output, 2),
+							 THCudaTensor_stride(state, output, 3),
+							 THCState_getCurrentStream(state));
+
+  //check for errors
+  if (!success) {
+    THError("aborting");
+  }
+  return 1;
+}
+
+int BilinearSamplerBHWD_updateGradInput_cuda(THCudaTensor *inputImages, THCudaTensor *grids, THCudaTensor *gradInputImages,
+                                        THCudaTensor *gradGrids, THCudaTensor *gradOutput)
+{
+  int success = 0;
+  success = BilinearSamplerBHWD_updateGradInput_cuda_kernel(THCudaTensor_size(state, gradOutput, 1),
+							    THCudaTensor_size(state, gradOutput, 3),
+							    THCudaTensor_size(state, gradOutput, 2),
+							    THCudaTensor_size(state, gradOutput, 0),
+							    THCudaTensor_size(state, inputImages, 1),
+							    THCudaTensor_size(state, inputImages, 2),
+							    THCudaTensor_size(state, inputImages, 3),
+							    THCudaTensor_size(state, inputImages, 0),
+							    THCudaTensor_data(state, inputImages),
+							    THCudaTensor_stride(state, inputImages, 0),
+							    THCudaTensor_stride(state, inputImages, 1),
+							    THCudaTensor_stride(state, inputImages, 2),
+							    THCudaTensor_stride(state, inputImages, 3),
+							    THCudaTensor_data(state, grids),
+							    THCudaTensor_stride(state, grids, 0),
+							    THCudaTensor_stride(state, grids, 3),
+							    THCudaTensor_stride(state, grids, 1),
+							    THCudaTensor_stride(state, grids, 2),
+							    THCudaTensor_data(state, gradInputImages),
+							    THCudaTensor_stride(state, gradInputImages, 0),
+							    THCudaTensor_stride(state, gradInputImages, 1),
+							    THCudaTensor_stride(state, gradInputImages, 2),
+							    THCudaTensor_stride(state, gradInputImages, 3),
+							    THCudaTensor_data(state, gradGrids),
+							    THCudaTensor_stride(state, gradGrids, 0),
+							    THCudaTensor_stride(state, gradGrids, 3),
+							    THCudaTensor_stride(state, gradGrids, 1),
+							    THCudaTensor_stride(state, gradGrids, 2),
+							    THCudaTensor_data(state, gradOutput),
+							    THCudaTensor_stride(state, gradOutput, 0),
+							    THCudaTensor_stride(state, gradOutput, 1),
+							    THCudaTensor_stride(state, gradOutput, 2),
+							    THCudaTensor_stride(state, gradOutput, 3),
+							    THCState_getCurrentStream(state));
+
+  //check for errors
+  if (!success) {
+    THError("aborting");
+  }
+  return 1;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward", &BilinearSamplerBHWD_updateOutput_cuda, "roi crop forward (CUDA)");
+  m.def("backward", &BilinearSamplerBHWD_updateGradInput_cuda, "roi crop backward (CUDA)");
+}
\ No newline at end of file
diff --git a/lib/ops/roi_crop/src/roi_crop_kernel.cu b/lib/ops/roi_crop/src/roi_crop_kernel.cu
new file mode 100644
index 0000000..dc00011
--- /dev/null
+++ b/lib/ops/roi_crop/src/roi_crop_kernel.cu
@@ -0,0 +1,251 @@
+#include <stdbool.h>
+#include <stdio.h>
+
+#define real float
+
+// Bilinear sampling is done in BHWD (coalescing is not obvious in BDHW)
+// we assume BHWD format in inputImages
+// we assume BHW(YX) format on grids
+
+__device__ void getTopLeft(float x, int width, int& point, float& weight)
+{
+   /* for interpolation :
+      stores in point and weight :
+      - the x-coordinate of the pixel on the left (or y-coordinate of the upper pixel)
+      - the weight for interpolating
+   */
+
+   float xcoord = (x + 1) * (width - 1) / 2;
+   point = floor(xcoord);
+   weight = 1 - (xcoord - point);
+}
+
+__device__ bool between(int value, int lowerBound, int upperBound)
+{
+   return (value >= lowerBound && value <= upperBound);
+}
+
+__device__ void sumReduceShMem(volatile float s[])
+{
+   /* obviously only works for 32 elements */
+   /* sums up a shared memory array of 32 elements, stores it in s[0] */
+   /* whole warp can then read first element (broadcasting) */
+   if(threadIdx.x<16) { s[threadIdx.x] = s[threadIdx.x] + s[threadIdx.x+16]; }
+   if(threadIdx.x<8) { s[threadIdx.x] = s[threadIdx.x] + s[threadIdx.x+8]; }
+   if(threadIdx.x<4) { s[threadIdx.x] = s[threadIdx.x] + s[threadIdx.x+4]; }
+   if(threadIdx.x<2) { s[threadIdx.x] = s[threadIdx.x] + s[threadIdx.x+2]; }
+   if(threadIdx.x<1) { s[threadIdx.x] = s[threadIdx.x] + s[threadIdx.x+1]; }
+}
+
+// CUDA: grid stride looping
+#define CUDA_KERNEL_LOOP(i, n) \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; \
+       i < (n); \
+       i += blockDim.x * gridDim.x)
+
+__global__ void bilinearSamplingFromGrid(const int nthreads, float* inputImages_data, int inputImages_strideBatch, int inputImages_strideChannels, int inputImages_strideHeight, int inputImages_strideWidth,
+                                         float* grids_data, int grids_strideBatch, int grids_strideYX, int grids_strideHeight, int grids_strideWidth,
+                                         float* output_data, int output_strideBatch, int output_strideChannels, int output_strideHeight, int output_strideWidth,
+                                         int inputImages_channels, int inputImages_height, int inputImages_width,
+                                         int output_channels, int output_height, int output_width, int output_batchsize,
+                                         int roiPerImage)
+{
+   CUDA_KERNEL_LOOP(index, nthreads)
+   {
+       const int xOut = index % output_width;
+       const int yOut = (index / output_width) % output_height;
+       const int cOut  = (index / output_width / output_height) % output_channels;
+       const int b = index / output_width / output_height / output_channels;
+
+       const int width = inputImages_width;
+       const int height = inputImages_height;
+
+       const int b_input = b / roiPerImage;
+
+       float yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+       float xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1];
+
+       int yInTopLeft, xInTopLeft;
+       float yWeightTopLeft, xWeightTopLeft;
+       getTopLeft(xf, inputImages_width, xInTopLeft, xWeightTopLeft);
+       getTopLeft(yf, inputImages_height, yInTopLeft, yWeightTopLeft);
+
+       // const int outAddress = output_strideBatch * b + output_strideHeight * yOut + output_strideWidth * xOut;
+       const int outAddress = output_strideBatch * b + output_strideChannels * cOut + output_strideHeight * yOut + xOut;
+
+       const int inTopLeftAddress = inputImages_strideBatch * b_input + inputImages_strideChannels * cOut + inputImages_strideHeight * yInTopLeft + xInTopLeft;
+       const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+       const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+       const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+       float v=0;
+       float inTopLeft=0;
+       float inTopRight=0;
+       float inBottomLeft=0;
+       float inBottomRight=0;
+
+       bool topLeftIsIn = between(xInTopLeft, 0, width-1) && between(yInTopLeft, 0, height-1);
+       bool topRightIsIn = between(xInTopLeft+1, 0, width-1) && between(yInTopLeft, 0, height-1);
+       bool bottomLeftIsIn = between(xInTopLeft, 0, width-1) && between(yInTopLeft+1, 0, height-1);
+       bool bottomRightIsIn = between(xInTopLeft+1, 0, width-1) && between(yInTopLeft+1, 0, height-1);
+
+       if (!topLeftIsIn && !topRightIsIn && !bottomLeftIsIn && !bottomRightIsIn)
+         continue;
+
+       if(topLeftIsIn) inTopLeft = inputImages_data[inTopLeftAddress];
+       if(topRightIsIn) inTopRight = inputImages_data[inTopRightAddress];
+       if(bottomLeftIsIn) inBottomLeft = inputImages_data[inBottomLeftAddress];
+       if(bottomRightIsIn) inBottomRight = inputImages_data[inBottomRightAddress];
+
+       v = xWeightTopLeft * yWeightTopLeft * inTopLeft
+         + (1 - xWeightTopLeft) * yWeightTopLeft * inTopRight
+         + xWeightTopLeft * (1 - yWeightTopLeft) * inBottomLeft
+         + (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * inBottomRight;
+
+       output_data[outAddress] = v;
+   }
+
+}
+
+__global__ void backwardBilinearSampling(const int nthreads, float* inputImages_data, int inputImages_strideBatch, int inputImages_strideChannels, int inputImages_strideHeight, int inputImages_strideWidth,
+                                         float* gradInputImages_data, int gradInputImages_strideBatch, int gradInputImages_strideChannels, int gradInputImages_strideHeight, int gradInputImages_strideWidth,
+                                         float* grids_data, int grids_strideBatch, int grids_strideYX, int grids_strideHeight, int grids_strideWidth,
+                                         float* gradGrids_data, int gradGrids_strideBatch, int gradGrids_strideYX, int gradGrids_strideHeight, int gradGrids_strideWidth,
+                                         float* gradOutput_data, int gradOutput_strideBatch, int gradOutput_strideChannels, int gradOutput_strideHeight, int gradOutput_strideWidth,
+                                         int inputImages_channels, int inputImages_height, int inputImages_width,
+                                         int gradOutput_channels, int gradOutput_height, int gradOutput_width, int gradOutput_batchsize,
+                                         int roiPerImage)
+{
+
+  CUDA_KERNEL_LOOP(index, nthreads)
+  {
+      const int xOut = index % gradOutput_width;
+      const int yOut = (index / gradOutput_width) % gradOutput_height;
+      const int cOut  = (index / gradOutput_width / gradOutput_height) % gradOutput_channels;
+      const int b = index / gradOutput_width / gradOutput_height / gradOutput_channels;
+
+      const int b_input = b / roiPerImage;
+
+      const int width = inputImages_width;
+      const int height = inputImages_height;
+
+      float yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
+      float xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1];
+
+      int yInTopLeft, xInTopLeft;
+      float yWeightTopLeft, xWeightTopLeft;
+      getTopLeft(xf, inputImages_width, xInTopLeft, xWeightTopLeft);
+      getTopLeft(yf, inputImages_height, yInTopLeft, yWeightTopLeft);
+
+      const int inTopLeftAddress = inputImages_strideBatch * b_input + inputImages_strideChannels * cOut + inputImages_strideHeight * yInTopLeft + xInTopLeft;
+      const int inTopRightAddress = inTopLeftAddress + inputImages_strideWidth;
+      const int inBottomLeftAddress = inTopLeftAddress + inputImages_strideHeight;
+      const int inBottomRightAddress = inBottomLeftAddress + inputImages_strideWidth;
+
+      const int gradInputImagesTopLeftAddress = gradInputImages_strideBatch * b_input + gradInputImages_strideChannels * cOut
+                                              + gradInputImages_strideHeight * yInTopLeft + xInTopLeft;
+      const int gradInputImagesTopRightAddress = gradInputImagesTopLeftAddress + gradInputImages_strideWidth;
+      const int gradInputImagesBottomLeftAddress = gradInputImagesTopLeftAddress + gradInputImages_strideHeight;
+      const int gradInputImagesBottomRightAddress = gradInputImagesBottomLeftAddress + gradInputImages_strideWidth;
+
+      const int gradOutputAddress = gradOutput_strideBatch * b + gradOutput_strideChannels * cOut + gradOutput_strideHeight * yOut + xOut;
+
+      float topLeftDotProduct = 0;
+      float topRightDotProduct = 0;
+      float bottomLeftDotProduct = 0;
+      float bottomRightDotProduct = 0;
+
+      bool topLeftIsIn = between(xInTopLeft, 0, width-1) && between(yInTopLeft, 0, height-1);
+      bool topRightIsIn = between(xInTopLeft+1, 0, width-1) && between(yInTopLeft, 0, height-1);
+      bool bottomLeftIsIn = between(xInTopLeft, 0, width-1) && between(yInTopLeft+1, 0, height-1);
+      bool bottomRightIsIn = between(xInTopLeft+1, 0, width-1) && between(yInTopLeft+1, 0, height-1);
+
+      float gradOutValue = gradOutput_data[gradOutputAddress];
+      // bool between(int value, int lowerBound, int upperBound)
+      if(topLeftIsIn)
+      {
+         float inTopLeft = inputImages_data[inTopLeftAddress];
+         topLeftDotProduct += inTopLeft * gradOutValue;
+         atomicAdd(&gradInputImages_data[gradInputImagesTopLeftAddress], xWeightTopLeft * yWeightTopLeft * gradOutValue);
+      }
+
+      if(topRightIsIn)
+      {
+         float inTopRight = inputImages_data[inTopRightAddress];
+         topRightDotProduct += inTopRight * gradOutValue;
+         atomicAdd(&gradInputImages_data[gradInputImagesTopRightAddress], (1 - xWeightTopLeft) * yWeightTopLeft * gradOutValue);
+      }
+
+      if(bottomLeftIsIn)
+      {
+         float inBottomLeft = inputImages_data[inBottomLeftAddress];
+         bottomLeftDotProduct += inBottomLeft * gradOutValue;
+         atomicAdd(&gradInputImages_data[gradInputImagesBottomLeftAddress], xWeightTopLeft * (1 - yWeightTopLeft) * gradOutValue);
+      }
+
+      if(bottomRightIsIn)
+      {
+         float inBottomRight = inputImages_data[inBottomRightAddress];
+         bottomRightDotProduct += inBottomRight * gradOutValue;
+         atomicAdd(&gradInputImages_data[gradInputImagesBottomRightAddress], (1 - xWeightTopLeft) * (1 - yWeightTopLeft) * gradOutValue);
+      }
+  }
+}
+
+
+
+
+int BilinearSamplerBHWD_updateOutput_cuda_kernel(int oc, int ow, int oh, int ob, int ic, int ih, int iw, int ib,
+                                                 float *inputImages, int isb, int isc, int ish, int isw,
+                                                 float *grids, int gsb, int gsc, int gsh, int gsw,
+                                                 float *output, int osb, int osc, int osh, int osw,
+                                                 cudaStream_t stream)
+{
+   const int kThreadsPerBlock = 1024;
+   int output_size = ob * oh * ow * oc;
+   cudaError_t err;
+   int roiPerImage = ob / ib;
+
+   bilinearSamplingFromGrid<<<(output_size + kThreadsPerBlock - 1) / kThreadsPerBlock, kThreadsPerBlock, 0, stream>>>(
+     output_size, inputImages, isb, isc, ish, isw, grids, gsb, gsc, gsh, gsw, output, osb, osc, osh, osw, ic, ih, iw, oc, oh, ow,
+     ob, roiPerImage);
+
+   // check for errors
+   err = cudaGetLastError();
+   if (err != cudaSuccess) {
+     printf("error in BilinearSampler.updateOutput: %s\n", cudaGetErrorString(err));
+     //THError("aborting");
+     return 0;
+   }
+   return 1;
+}
+
+int BilinearSamplerBHWD_updateGradInput_cuda_kernel(int goc, int gow, int goh, int gob, int ic, int ih, int iw, int ib,
+                                                    float *inputImages, int isb, int isc, int ish, int isw,
+                                                    float *grids, int gsb, int gsc, int gsh, int gsw,
+                                                    float *gradInputImages, int gisb, int gisc, int gish, int gisw,
+                                                    float *gradGrids, int ggsb, int ggsc, int ggsh, int ggsw,
+                                                    float *gradOutput, int gosb, int gosc, int gosh, int gosw,
+                                                    cudaStream_t stream)
+{
+
+  const int kThreadsPerBlock = 1024;
+  int output_size = gob * goh * gow * goc;
+  cudaError_t err;
+  int roiPerImage = gob / ib;
+
+
+  backwardBilinearSampling<<<(output_size + kThreadsPerBlock - 1) / kThreadsPerBlock, kThreadsPerBlock, 0, stream>>>(
+    output_size, inputImages, isb, isc, ish, isw, gradInputImages, gisb, gisc, gish, gisw, grids, gsb, gsc, gsh, gsw,
+    gradGrids, ggsb, ggsc, ggsh, ggsw, gradOutput, gosb, gosc, gosh, gosw, ic, ih, iw, goc, goh, gow, gob,
+    roiPerImage);
+
+  // check for errors
+  err = cudaGetLastError();
+  if (err != cudaSuccess) {
+    printf("error in BilinearSampler.updateGradInput: %s\n", cudaGetErrorString(err));
+    //THError("aborting");
+    return 0;
+  }
+  return 1;
+}
\ No newline at end of file
diff --git a/lib/ops/roi_pool/__init__.py b/lib/ops/roi_pool/__init__.py
new file mode 100644
index 0000000..9f0474e
--- /dev/null
+++ b/lib/ops/roi_pool/__init__.py
@@ -0,0 +1,3 @@
+from .roi_pool import RoIPool, roi_pool
+
+__all__ = ['roi_pool', 'RoIPool']
diff --git a/lib/ops/roi_pool/__pycache__/__init__.cpython-37.pyc b/lib/ops/roi_pool/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..a5bc310
Binary files /dev/null and b/lib/ops/roi_pool/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/roi_pool/__pycache__/roi_pool.cpython-37.pyc b/lib/ops/roi_pool/__pycache__/roi_pool.cpython-37.pyc
new file mode 100644
index 0000000..ee12acc
Binary files /dev/null and b/lib/ops/roi_pool/__pycache__/roi_pool.cpython-37.pyc differ
diff --git a/lib/ops/roi_pool/gradcheck.py b/lib/ops/roi_pool/gradcheck.py
new file mode 100644
index 0000000..d11af79
--- /dev/null
+++ b/lib/ops/roi_pool/gradcheck.py
@@ -0,0 +1,16 @@
+import os.path as osp
+import sys
+
+import torch
+from torch.autograd import gradcheck
+
+sys.path.append(osp.abspath(osp.join(__file__, '../../')))
+from roi_pool import RoIPool  # noqa: E402, isort:skip
+
+feat = torch.randn(4, 16, 15, 15, requires_grad=True).cuda()
+rois = torch.Tensor([[0, 0, 0, 50, 50], [0, 10, 30, 43, 55],
+                     [1, 67, 40, 110, 120]]).cuda()
+inputs = (feat, rois)
+print('Gradcheck for roi pooling...')
+test = gradcheck(RoIPool(4, 1.0 / 8), inputs, eps=1e-5, atol=1e-3)
+print(test)
diff --git a/lib/ops/roi_pool/roi_pool.py b/lib/ops/roi_pool/roi_pool.py
new file mode 100644
index 0000000..26d900f
--- /dev/null
+++ b/lib/ops/roi_pool/roi_pool.py
@@ -0,0 +1,75 @@
+import torch
+import torch.nn as nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.modules.utils import _pair
+
+from . import roi_pool_cuda
+
+
+class RoIPoolFunction(Function):
+
+    @staticmethod
+    def forward(ctx, features, rois, out_size, spatial_scale):
+        assert features.is_cuda
+        out_h, out_w = _pair(out_size)
+        assert isinstance(out_h, int) and isinstance(out_w, int)
+        ctx.save_for_backward(rois)
+        num_channels = features.size(1)
+        num_rois = rois.size(0)
+        out_size = (num_rois, num_channels, out_h, out_w)
+        output = features.new_zeros(out_size)
+        argmax = features.new_zeros(out_size, dtype=torch.int)
+        roi_pool_cuda.forward(features, rois, out_h, out_w, spatial_scale,
+                              output, argmax)
+        ctx.spatial_scale = spatial_scale
+        ctx.feature_size = features.size()
+        ctx.argmax = argmax
+
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        assert grad_output.is_cuda
+        spatial_scale = ctx.spatial_scale
+        feature_size = ctx.feature_size
+        argmax = ctx.argmax
+        rois = ctx.saved_tensors[0]
+        assert feature_size is not None
+
+        grad_input = grad_rois = None
+        if ctx.needs_input_grad[0]:
+            grad_input = grad_output.new_zeros(feature_size)
+            roi_pool_cuda.backward(grad_output.contiguous(), rois, argmax,
+                                   spatial_scale, grad_input)
+
+        return grad_input, grad_rois, None, None
+
+
+roi_pool = RoIPoolFunction.apply
+
+
+class RoIPool(nn.Module):
+
+    def __init__(self, out_size, spatial_scale, use_torchvision=False):
+        super(RoIPool, self).__init__()
+
+        self.out_size = _pair(out_size)
+        self.spatial_scale = float(spatial_scale)
+        self.use_torchvision = use_torchvision
+
+    def forward(self, features, rois):
+        if self.use_torchvision:
+            from torchvision.ops import roi_pool as tv_roi_pool
+            return tv_roi_pool(features, rois, self.out_size,
+                               self.spatial_scale)
+        else:
+            return roi_pool(features, rois, self.out_size, self.spatial_scale)
+
+    def __repr__(self):
+        format_str = self.__class__.__name__
+        format_str += '(out_size={}, spatial_scale={}'.format(
+            self.out_size, self.spatial_scale)
+        format_str += ', use_torchvision={})'.format(self.use_torchvision)
+        return format_str
diff --git a/lib/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..4b5057f
Binary files /dev/null and b/lib/ops/roi_pool/roi_pool_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/roi_pool/src/roi_pool_cuda.cpp b/lib/ops/roi_pool/src/roi_pool_cuda.cpp
new file mode 100644
index 0000000..84c4e4b
--- /dev/null
+++ b/lib/ops/roi_pool/src/roi_pool_cuda.cpp
@@ -0,0 +1,86 @@
+#include <torch/extension.h>
+
+#include <cmath>
+#include <vector>
+
+int ROIPoolForwardLaucher(const at::Tensor features, const at::Tensor rois,
+                          const float spatial_scale, const int channels,
+                          const int height, const int width, const int num_rois,
+                          const int pooled_h, const int pooled_w,
+                          at::Tensor output, at::Tensor argmax);
+
+int ROIPoolBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
+                           const at::Tensor argmax, const float spatial_scale,
+                           const int batch_size, const int channels,
+                           const int height, const int width,
+                           const int num_rois, const int pooled_h,
+                           const int pooled_w, at::Tensor bottom_grad);
+
+#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
+#define CHECK_CONTIGUOUS(x) \
+  AT_CHECK(x.is_contiguous(), #x, " must be contiguous ")
+#define CHECK_INPUT(x) \
+  CHECK_CUDA(x);       \
+  CHECK_CONTIGUOUS(x)
+
+int roi_pooling_forward_cuda(at::Tensor features, at::Tensor rois,
+                             int pooled_height, int pooled_width,
+                             float spatial_scale, at::Tensor output,
+                             at::Tensor argmax) {
+  CHECK_INPUT(features);
+  CHECK_INPUT(rois);
+  CHECK_INPUT(output);
+  CHECK_INPUT(argmax);
+
+  // Number of ROIs
+  int num_rois = rois.size(0);
+  int size_rois = rois.size(1);
+
+  if (size_rois != 5) {
+    printf("wrong roi size\n");
+    return 0;
+  }
+
+  int channels = features.size(1);
+  int height = features.size(2);
+  int width = features.size(3);
+
+  ROIPoolForwardLaucher(features, rois, spatial_scale, channels, height, width,
+                        num_rois, pooled_height, pooled_width, output, argmax);
+
+  return 1;
+}
+
+int roi_pooling_backward_cuda(at::Tensor top_grad, at::Tensor rois,
+                              at::Tensor argmax, float spatial_scale,
+                              at::Tensor bottom_grad) {
+  CHECK_INPUT(top_grad);
+  CHECK_INPUT(rois);
+  CHECK_INPUT(argmax);
+  CHECK_INPUT(bottom_grad);
+
+  int pooled_height = top_grad.size(2);
+  int pooled_width = top_grad.size(3);
+  int num_rois = rois.size(0);
+  int size_rois = rois.size(1);
+
+  if (size_rois != 5) {
+    printf("wrong roi size\n");
+    return 0;
+  }
+  int batch_size = bottom_grad.size(0);
+  int channels = bottom_grad.size(1);
+  int height = bottom_grad.size(2);
+  int width = bottom_grad.size(3);
+
+  ROIPoolBackwardLaucher(top_grad, rois, argmax, spatial_scale, batch_size,
+                         channels, height, width, num_rois, pooled_height,
+                         pooled_width, bottom_grad);
+
+  return 1;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward", &roi_pooling_forward_cuda, "Roi_Pooling forward (CUDA)");
+  m.def("backward", &roi_pooling_backward_cuda, "Roi_Pooling backward (CUDA)");
+}
diff --git a/lib/ops/roi_pool/src/roi_pool_kernel.cu b/lib/ops/roi_pool/src/roi_pool_kernel.cu
new file mode 100644
index 0000000..67e15de
--- /dev/null
+++ b/lib/ops/roi_pool/src/roi_pool_kernel.cu
@@ -0,0 +1,156 @@
+#include <ATen/ATen.h>
+#include <THC/THCAtomics.cuh>
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                            \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
+       i += blockDim.x * gridDim.x)
+
+#define THREADS_PER_BLOCK 1024
+
+inline int GET_BLOCKS(const int N) {
+  int optimal_block_num = (N + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+  int max_block_num = 65000;
+  return min(optimal_block_num, max_block_num);
+}
+
+template <typename scalar_t>
+__global__ void ROIPoolForward(const int nthreads, const scalar_t *bottom_data,
+                               const scalar_t *rois,
+                               const scalar_t spatial_scale, const int channels,
+                               const int height, const int width,
+                               const int pooled_h, const int pooled_w,
+                               scalar_t *top_data, int *argmax_data) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    // (n, c, ph, pw) is an element in the pooled output
+    int pw = index % pooled_w;
+    int ph = (index / pooled_w) % pooled_h;
+    int c = (index / pooled_w / pooled_h) % channels;
+    int n = index / pooled_w / pooled_h / channels;
+
+    const scalar_t *offset_rois = rois + n * 5;
+    int roi_batch_ind = offset_rois[0];
+    // calculate the roi region on feature maps
+    scalar_t roi_x1 = offset_rois[1] * spatial_scale;
+    scalar_t roi_y1 = offset_rois[2] * spatial_scale;
+    scalar_t roi_x2 = (offset_rois[3] + 1) * spatial_scale;
+    scalar_t roi_y2 = (offset_rois[4] + 1) * spatial_scale;
+
+    // force malformed rois to be 1x1
+    scalar_t roi_w = roi_x2 - roi_x1;
+    scalar_t roi_h = roi_y2 - roi_y1;
+    if (roi_w <= 0 || roi_h <= 0) continue;
+
+    scalar_t bin_size_w = roi_w / static_cast<scalar_t>(pooled_w);
+    scalar_t bin_size_h = roi_h / static_cast<scalar_t>(pooled_h);
+
+    // the corresponding bin region
+    int bin_x1 = floor(static_cast<scalar_t>(pw) * bin_size_w + roi_x1);
+    int bin_y1 = floor(static_cast<scalar_t>(ph) * bin_size_h + roi_y1);
+    int bin_x2 = ceil(static_cast<scalar_t>(pw + 1) * bin_size_w + roi_x1);
+    int bin_y2 = ceil(static_cast<scalar_t>(ph + 1) * bin_size_h + roi_y1);
+
+    // add roi offsets and clip to input boundaries
+    bin_x1 = min(max(bin_x1, 0), width);
+    bin_y1 = min(max(bin_y1, 0), height);
+    bin_x2 = min(max(bin_x2, 0), width);
+    bin_y2 = min(max(bin_y2, 0), height);
+    bool is_empty = (bin_y2 <= bin_y1) || (bin_x2 <= bin_x1);
+
+    // If nothing is pooled, argmax = -1 causes nothing to be backprop'd
+    int max_idx = -1;
+    bottom_data += (roi_batch_ind * channels + c) * height * width;
+
+    // Define an empty pooling region to be zero
+    scalar_t max_val = is_empty ? static_cast<scalar_t>(0)
+                                : bottom_data[bin_y1 * width + bin_x1] - 1;
+
+    for (int h = bin_y1; h < bin_y2; ++h) {
+      for (int w = bin_x1; w < bin_x2; ++w) {
+        int offset = h * width + w;
+        if (bottom_data[offset] > max_val) {
+          max_val = bottom_data[offset];
+          max_idx = offset;
+        }
+      }
+    }
+    top_data[index] = max_val;
+    if (argmax_data != NULL) argmax_data[index] = max_idx;
+  }
+}
+
+int ROIPoolForwardLaucher(const at::Tensor features, const at::Tensor rois,
+                          const float spatial_scale, const int channels,
+                          const int height, const int width, const int num_rois,
+                          const int pooled_h, const int pooled_w,
+                          at::Tensor output, at::Tensor argmax) {
+const int output_size = num_rois * channels * pooled_h * pooled_w;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      features.scalar_type(), "ROIPoolLaucherForward", ([&] {
+        const scalar_t *bottom_data = features.data<scalar_t>();
+        const scalar_t *rois_data = rois.data<scalar_t>();
+        scalar_t *top_data = output.data<scalar_t>();
+        int *argmax_data = argmax.data<int>();
+
+        ROIPoolForward<scalar_t>
+            <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK>>>(
+                output_size, bottom_data, rois_data, scalar_t(spatial_scale),
+                channels, height, width, pooled_h, pooled_w, top_data,
+                argmax_data);
+      }));
+  THCudaCheck(cudaGetLastError());
+  return 1;
+}
+
+template <typename scalar_t>
+__global__ void ROIPoolBackward(const int nthreads, const scalar_t *top_diff,
+                                const scalar_t *rois, const int *argmax_data,
+                                const scalar_t spatial_scale,
+                                const int channels, const int height,
+                                const int width, const int pooled_h,
+                                const int pooled_w, scalar_t *bottom_diff) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    int pw = index % pooled_w;
+    int ph = (index / pooled_w) % pooled_h;
+    int c = (index / pooled_w / pooled_h) % channels;
+    int n = index / pooled_w / pooled_h / channels;
+
+    int roi_batch_ind = rois[n * 5];
+    int bottom_index = argmax_data[(n * channels + c) * pooled_h * pooled_w +
+                                   ph * pooled_w + pw];
+
+    atomicAdd(bottom_diff + (roi_batch_ind * channels + c) * height * width +
+                  bottom_index,
+              top_diff[index]);
+  }
+}
+
+int ROIPoolBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
+                           const at::Tensor argmax, const float spatial_scale,
+                           const int batch_size, const int channels,
+                           const int height, const int width,
+                           const int num_rois, const int pooled_h,
+                           const int pooled_w, at::Tensor bottom_grad) {
+  const int output_size = num_rois * pooled_h * pooled_w * channels;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      top_grad.scalar_type(), "ROIPoolLaucherBackward", ([&] {
+        const scalar_t *top_diff = top_grad.data<scalar_t>();
+        const scalar_t *rois_data = rois.data<scalar_t>();
+        const int *argmax_data = argmax.data<int>();
+        scalar_t *bottom_diff = bottom_grad.data<scalar_t>();
+
+        if (sizeof(scalar_t) == sizeof(double)) {
+          fprintf(stderr, "double is not supported\n");
+          exit(-1);
+        }
+
+        ROIPoolBackward<scalar_t>
+            <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK>>>(
+                output_size, top_diff, rois_data, argmax_data,
+                scalar_t(spatial_scale), channels, height, width, pooled_h,
+                pooled_w, bottom_diff);
+      }));
+  THCudaCheck(cudaGetLastError());
+  return 1;
+}
diff --git a/lib/ops/roi_ring_pool/__init__.py b/lib/ops/roi_ring_pool/__init__.py
new file mode 100644
index 0000000..944bd3d
--- /dev/null
+++ b/lib/ops/roi_ring_pool/__init__.py
@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 29 15:18:12 2019
+
+@author: vasgaoweithu
+"""
+
+from .roi_ring_pool import RoIRingPool
+
+__all__ = ['RoIRingPool']
\ No newline at end of file
diff --git a/lib/ops/roi_ring_pool/__pycache__/__init__.cpython-37.pyc b/lib/ops/roi_ring_pool/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..f3025b2
Binary files /dev/null and b/lib/ops/roi_ring_pool/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/ops/roi_ring_pool/__pycache__/roi_ring_pool.cpython-37.pyc b/lib/ops/roi_ring_pool/__pycache__/roi_ring_pool.cpython-37.pyc
new file mode 100644
index 0000000..178c01e
Binary files /dev/null and b/lib/ops/roi_ring_pool/__pycache__/roi_ring_pool.cpython-37.pyc differ
diff --git a/lib/ops/roi_ring_pool/roi_ring_pool.py b/lib/ops/roi_ring_pool/roi_ring_pool.py
new file mode 100644
index 0000000..2719925
--- /dev/null
+++ b/lib/ops/roi_ring_pool/roi_ring_pool.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 29 14:50:56 2019
+
+@author: vasgaoweithu
+"""
+
+import torch
+import torch.nn as nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from . import roi_ring_pool_cuda
+
+class RoIRingPoolFunction(Function):
+    @staticmethod
+    def forward(ctx, pooled_height, pooled_width, spatial_scale, scale_inner, scale_outer, features, rois):
+        ctx.pooled_height = pooled_height
+        ctx.pooled_width = pooled_width
+        ctx.spatial_scale = spatial_scale
+        ctx.scale_inner = scale_inner
+        ctx.scale_outer = scale_outer
+        ctx.feature_size = features.size()
+        batch_size, num_channels, data_height, data_width = ctx.feature_size
+        num_rois = rois.size(0)
+        output = features.new(num_rois, num_channels, ctx.pooled_height, ctx.pooled_width).zero_()
+        ctx.argmax = features.new(num_rois, num_channels, ctx.pooled_height, ctx.pooled_width).zero_().int()
+
+        ctx.rois = rois
+        ctx.processed_rois = features.new(rois.size(0), 9).zero_()
+        
+        RectangularRing(rois, ctx.processed_rois, ctx.spatial_scale, ctx.scale_inner, ctx.scale_outer)
+        roi_ring_pool_cuda.forward(ctx.pooled_height, ctx.pooled_width, ctx.spatial_scale,
+                                                           features, ctx.processed_rois, output, ctx.argmax)
+        return output
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        assert(ctx.feature_size is not None and grad_output.is_cuda)
+        batch_size, num_channels, data_height, data_width = ctx.feature_size
+        grad_input = grad_output.new(batch_size, num_channels, data_height, data_width).zero_()
+
+        roi_ring_pool_cuda.backward(ctx.pooled_height, ctx.pooled_width, ctx.spatial_scale,
+                                                        grad_output, ctx.processed_rois, grad_input, ctx.argmax)
+        return None, None, None, None, None, grad_input, None
+    
+roi_ring_pool = RoIRingPoolFunction.apply
+
+class RoIRingPool(nn.Module):
+    def __init__(self, pooled_height, pooled_width, spatial_scale,  scale_inner, scale_outer):
+        super(RoIRingPool, self).__init__()
+        
+        self.pooled_width = int(pooled_width)
+        self.pooled_height = int(pooled_height)
+        self.spatial_scale = float(spatial_scale)
+        self.scale_inner = scale_inner
+        self.scale_outer = scale_outer
+    def forward(self, features, rois):
+        return roi_ring_pool(self.pooled_height, self.pooled_width, self.spatial_scale, self.scale_inner, self.scale_outer,
+                                   features, rois)
+
+
+def RectangularRing(ss_rois, processed_rois,spatial_scale, scale_inner, scale_outer):
+    #widths = rois[:, 3] - rois[:, 1] + 1.0
+    #heights = rois[:, 4] - rois[:, 2] + 1.0
+    #ctr_x = rois[:, 1] + 0.5 * widths
+    #ctr_y = rois[:, 2] + 0.5 * heights
+    
+    rois = ss_rois.clone()
+
+    ctr_x = (rois[:, 1] + rois[:, 3]) / 2
+    ctr_y = (rois[:, 2] + rois[:, 4]) / 2
+    w_half = (rois[:, 3] - rois[:, 1]) / 2
+    h_half = (rois[:, 4] - rois[:, 2]) / 2
+    
+    # processed_rois[:, 1] = torch.tensor(ctr_x - w_half * scale_outer, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(0.5).floor_()
+    # processed_rois[:, 2] = torch.tensor(ctr_y - h_half * scale_outer, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(0.5).floor_()
+    # processed_rois[:, 3] = torch.tensor(ctr_x + w_half * scale_outer, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(-0.5).ceil_()
+    # processed_rois[:, 4] = torch.tensor(ctr_y + h_half * scale_outer, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(-0.5).ceil_()
+    # processed_rois[:, 5] = torch.tensor(ctr_x - w_half * scale_inner, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(0.5).floor_()
+    # processed_rois[:, 6] = torch.tensor(ctr_y - h_half * scale_inner, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(0.5).floor_()
+    # processed_rois[:, 7] = torch.tensor(ctr_x + w_half * scale_inner, dtype=rois.dtype, device=rois.device)  ##.mul_(spatial_scale).add_(-0.5).ceil_()
+    # processed_rois[:, 8] = torch.tensor(ctr_y + h_half * scale_inner, dtype=rois.dtype, device=rois.device)
+
+    processed_rois[:, 1] = (ctr_x - w_half * scale_outer).clone().detach().float().to(rois.device)
+    processed_rois[:, 2] = (ctr_y - h_half * scale_outer).clone().detach().float().to(rois.device)
+    processed_rois[:, 3] = (ctr_x + w_half * scale_outer).clone().detach().float().to(rois.device)
+    processed_rois[:, 4] = (ctr_y + h_half * scale_outer).clone().detach().float().to(rois.device)
+    processed_rois[:, 5] = (ctr_x - w_half * scale_inner).clone().detach().float().to(rois.device)
+    processed_rois[:, 6] = (ctr_y - h_half * scale_inner).clone().detach().float().to(rois.device)
+    processed_rois[:, 7] = (ctr_x + w_half * scale_inner).clone().detach().float().to(rois.device)
+    processed_rois[:, 8] = (ctr_y + h_half * scale_inner).clone().detach().float().to(rois.device)
+
+    if scale_inner == 0:
+        processed_rois[:, 5:] = 0
+
+    return 1
\ No newline at end of file
diff --git a/lib/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so b/lib/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so
new file mode 100644
index 0000000..69ee1a3
Binary files /dev/null and b/lib/ops/roi_ring_pool/roi_ring_pool_cuda.cpython-37m-x86_64-linux-gnu.so differ
diff --git a/lib/ops/roi_ring_pool/src/roi_ring_pool_cuda.cpp b/lib/ops/roi_ring_pool/src/roi_ring_pool_cuda.cpp
new file mode 100644
index 0000000..8e46969
--- /dev/null
+++ b/lib/ops/roi_ring_pool/src/roi_ring_pool_cuda.cpp
@@ -0,0 +1,87 @@
+#include <torch/extension.h>
+
+#include <cmath>
+#include <vector>
+
+int ROIRingPoolForwardLaucher(
+    const at::Tensor bottom_data, const float spatial_scale, const int num_rois, const int height,
+    const int width, const int channels, const int pooled_height,
+    const int pooled_width, const at::Tensor bottom_rois,
+    at::Tensor top_data, at::Tensor argmax_data);
+
+int ROIRingPoolBackwardLaucher(const at::Tensor top_diff, const float spatial_scale, const int batch_size, const int num_rois,
+    const int height, const int width, const int channels, const int pooled_height,
+    const int pooled_width, const at::Tensor bottom_rois,
+    at::Tensor bottom_diff, const at::Tensor argmax_data);
+
+#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
+#define CHECK_CONTIGUOUS(x) \
+  AT_CHECK(x.is_contiguous(), #x, " must be contiguous ")
+#define CHECK_INPUT(x) \
+  CHECK_CUDA(x);       \
+  CHECK_CONTIGUOUS(x)
+
+int roi_ring_pool_forward_cuda(int pooled_height, int pooled_width, float spatial_scale,
+                               at::Tensor features, at::Tensor rois, at::Tensor output, at::Tensor argmax)
+{
+    // Grab the input tensor
+    CHECK_INPUT(features);
+    CHECK_INPUT(rois);
+    CHECK_INPUT(output);
+    CHECK_INPUT(argmax);
+
+    // Number of ROIs
+    int num_rois = rois.size(0);
+    int size_rois = rois.size(1);
+    if (size_rois != 9)
+    {
+        return 0;
+    }
+    // data height
+    int data_height = features.size(2);
+    // data width
+    int data_width = features.size(3);
+    // Number of channels
+    int num_channels = features.size(1);
+
+    ROIRingPoolForwardLaucher(
+        features, spatial_scale, num_rois, data_height,
+        data_width, num_channels, pooled_height,
+        pooled_width, rois,
+        output, argmax);
+
+    return 1;
+}
+
+int roi_ring_pool_backward_cuda(int pooled_height, int pooled_width, float spatial_scale,
+                                at::Tensor top_grad, at::Tensor rois, at::Tensor bottom_grad, at::Tensor argmax)
+{
+    // Number of ROIs
+    int num_rois = rois.size(0);
+    int size_rois = rois.size(1);
+    if (size_rois != 9)
+    {
+        return 0;
+    }
+
+    // batch size
+    int batch_size = bottom_grad.size(0);
+    // data height
+    int data_height = bottom_grad.size(2);
+    // data width
+    int data_width = bottom_grad.size(3);
+    // Number of channels
+    int num_channels = bottom_grad.size(1);
+    ROIRingPoolBackwardLaucher(
+        top_grad, spatial_scale, batch_size, num_rois, data_height,
+        data_width, num_channels, pooled_height,
+        pooled_width, rois,
+        bottom_grad, argmax);
+
+    return 1;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward", &roi_ring_pool_forward_cuda, "Roi_Ring_Pooling forward (CUDA)");
+  m.def("backward", &roi_ring_pool_backward_cuda, "Roi_Ring_Pooling backward (CUDA)");
+}
\ No newline at end of file
diff --git a/lib/ops/roi_ring_pool/src/roi_ring_pool_kernel.cu b/lib/ops/roi_ring_pool/src/roi_ring_pool_kernel.cu
new file mode 100644
index 0000000..4531eae
--- /dev/null
+++ b/lib/ops/roi_ring_pool/src/roi_ring_pool_kernel.cu
@@ -0,0 +1,219 @@
+#include <ATen/ATen.h>
+#include <THC/THCAtomics.cuh>
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                            \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
+       i += blockDim.x * gridDim.x)
+#define CUDA_KERNEL_LOOP(i, n) \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; \
+       i < (n); \
+       i += blockDim.x * gridDim.x)
+
+#define THREADS_PER_BLOCK 1024
+
+template <typename scalar_t>
+__global__ void ROIRingPoolForward(const int nthreads, const scalar_t* bottom_data,
+    const scalar_t spatial_scale, const int height, const int width, 
+    const int channels, const int pooled_height, const int pooled_width,
+    const scalar_t* bottom_rois, scalar_t* top_data, int* argmax_data) {
+	CUDA_KERNEL_LOOP(index, nthreads) {
+		// (n, c, ph, pw) is an element in the pooled output
+		int pw = index % pooled_width;
+		int ph = (index / pooled_width) % pooled_height;
+		int c = (index / pooled_width / pooled_height) % channels;
+		int n = index / pooled_width / pooled_height / channels;
+
+		// For each ROI R = [batch_index, x_outer_1, y_outer_1, x_outer_2, y_outer_2, x_inner_1, y_inner_1, x_inner_2, y_inner_2]: 
+		// where R_outer = [x_outer_1, y_outer_1, x_outer_2, y_outer_2] is the outer rectangle of the region and 
+		// R_inner = [x_inner_1, y_inner_1, x_inner_2, y_inner_2] is the inner rectangle of the region
+		// max pooler over R by ignoring (setting to zero) the activations that lay inside the inner rectangle R_inner
+
+		bottom_rois += n * 9;
+		int roi_batch_ind = bottom_rois[0];
+
+
+		// outer rectangle of the region
+		int roi_start_w   = int(bottom_rois[1] * spatial_scale);//* spatial_scale);
+		int roi_start_h   = int(bottom_rois[2] * spatial_scale);//* spatial_scale);
+		int roi_end_w     = int(bottom_rois[3] * spatial_scale);//* spatial_scale);
+		int roi_end_h     = int(bottom_rois[4] * spatial_scale);//* spatial_scale);
+
+		// inner rectangle of the region
+		int roi_start_w_in = int(bottom_rois[5] * spatial_scale);//* spatial_scale);
+		int roi_start_h_in = int(bottom_rois[6] * spatial_scale);//* spatial_scale);
+		int roi_end_w_in   = int(bottom_rois[7] * spatial_scale);//* spatial_scale);
+		int roi_end_h_in   = int(bottom_rois[8] * spatial_scale);//* spatial_scale);
+
+		// Force malformed ROIs to be 1x1
+		int roi_width  = max(roi_end_w - roi_start_w + 1, 1);
+		int roi_height = max(roi_end_h - roi_start_h + 1, 1);
+		scalar_t bin_size_h = static_cast<scalar_t>(roi_height) / static_cast<scalar_t>(pooled_height);
+		scalar_t bin_size_w = static_cast<scalar_t>(roi_width)  / static_cast<scalar_t>(pooled_width);
+
+		const int hstart = min(height, max(0, static_cast<int>(floor(static_cast<scalar_t>(ph)   * bin_size_h)) + roi_start_h));
+		const int hend   = min(height, max(0, static_cast<int>(ceil( static_cast<scalar_t>(ph+1) * bin_size_h)) + roi_start_h));
+		const int wstart = min(width,  max(0, static_cast<int>(floor(static_cast<scalar_t>(pw)   * bin_size_w)) + roi_start_w));
+		const int wend   = min(width,  max(0, static_cast<int>(ceil( static_cast<scalar_t>(pw+1) * bin_size_w)) + roi_start_w));
+
+		scalar_t maxval = 0; 
+
+		int maxidx = -1;
+		bottom_data += (roi_batch_ind * channels + c) * height * width;
+		for (int h = hstart; h < hend; ++h) {
+			for (int w = wstart; w < wend; ++w) {
+				if (!(w > roi_start_w_in && w < roi_end_w_in && h > roi_start_h_in && h < roi_end_h_in)) {
+					// if it is not inside the inner rectangle of the region
+					int bottom_index = h * width + w;
+					if (bottom_data[bottom_index] > maxval) {
+						maxval = bottom_data[bottom_index];
+						maxidx = bottom_index;
+					}
+				}
+			}
+		}
+		top_data[index] = maxval;
+		argmax_data[index] = maxidx;
+	}
+}
+
+int ROIRingPoolForwardLaucher(
+    const at::Tensor bottom_data, const float spatial_scale, const int num_rois, const int height,
+    const int width, const int channels, const int pooled_height,
+    const int pooled_width, const at::Tensor bottom_rois,
+    at::Tensor top_data, at::Tensor argmax_data)
+{
+    const int kThreadsPerBlock = 1024;
+    int output_size = num_rois * pooled_height * pooled_width * channels;
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+        bottom_data.scalar_type(), "ROIRingPoolForawrd", ([&]{
+            ROIRingPoolForward<scalar_t><<<(output_size + kThreadsPerBlock - 1) / kThreadsPerBlock, kThreadsPerBlock>>>(
+                                output_size, 
+                                bottom_data.data<scalar_t>(), 
+                                scalar_t(spatial_scale), 
+                                height, 
+                                width, 
+                                channels, 
+                                pooled_height,
+                                pooled_width, 
+                                bottom_rois.data<scalar_t>(), 
+                                top_data.data<scalar_t>(), 
+                                argmax_data.data<int>());
+        })
+    );
+
+    THCudaCheck(cudaGetLastError());
+    return 1;
+}
+
+template <typename scalar_t>
+__global__ void ROIRingPoolBackward(const int nthreads, const scalar_t* top_diff,
+    const int* argmax_data, const int num_rois, const scalar_t spatial_scale,
+    const int height, const int width, const int channels, 
+    const int pooled_height, const int pooled_width, scalar_t* bottom_diff,
+    const scalar_t* bottom_rois) {
+	CUDA_KERNEL_LOOP(index, nthreads) {
+		// (n, c, h, w) coords in bottom data
+		int w = index % width;
+		int h = (index / width) % height;
+		int c = (index / width / height) % channels;
+		int n = index / width / height / channels;
+
+		scalar_t gradient = 0;
+		// Accumulate gradient over all ROIs that pooled this element
+		for (int roi_n = 0; roi_n < num_rois; ++roi_n) {
+			const scalar_t* offset_bottom_rois = bottom_rois + roi_n * 9;
+			int roi_batch_ind = offset_bottom_rois[0];
+			// Skip if ROI's batch index doesn't match n
+			if (n != roi_batch_ind) {
+				continue;
+			}
+
+
+			// outer rectangle of the region
+			int roi_start_w   = int(offset_bottom_rois[1] * spatial_scale);// * spatial_scale);
+			int roi_start_h   = int(offset_bottom_rois[2] * spatial_scale);// * spatial_scale);
+			int roi_end_w     = int(offset_bottom_rois[3] * spatial_scale);// * spatial_scale);
+			int roi_end_h     = int(offset_bottom_rois[4] * spatial_scale);// * spatial_scale);
+
+			// inner rectangle of the region
+			int roi_start_w_in= int(offset_bottom_rois[5] * spatial_scale);// * spatial_scale);
+			int roi_start_h_in= int(offset_bottom_rois[6] * spatial_scale);// * spatial_scale);
+			int roi_end_w_in  = int(offset_bottom_rois[7] * spatial_scale);// * spatial_scale);
+			int roi_end_h_in  = int(offset_bottom_rois[8] * spatial_scale);// * spatial_scale);
+
+
+			// Skip if ROI doesn't include (h, w)
+			const bool in_roi =  (w >= roi_start_w && w <= roi_end_w &&
+					h >= roi_start_h && h <= roi_end_h) && 
+				!(w > roi_start_w_in && w < roi_end_w_in && 
+						h > roi_start_h_in && h < roi_end_h_in);
+
+			if (!in_roi) {
+				continue;
+			}
+
+			int top_offset = (roi_n * channels + c) * pooled_height * pooled_width;
+			const scalar_t* offset_top_diff = top_diff + top_offset;
+			const int* offset_argmax_data = argmax_data + top_offset;
+
+			// Compute feasible set of pooled units that could have pooled
+			// this bottom unit
+
+			// Force malformed ROIs to be 1x1
+			int roi_width = max(roi_end_w - roi_start_w + 1, 1);
+			int roi_height = max(roi_end_h - roi_start_h + 1, 1);
+
+			scalar_t bin_size_h = static_cast<scalar_t>(roi_height) / static_cast<scalar_t>(pooled_height);
+			scalar_t bin_size_w = static_cast<scalar_t>(roi_width)  / static_cast<scalar_t>(pooled_width);
+
+			int phstart = floor(static_cast<scalar_t>(h - roi_start_h) / bin_size_h);
+			int phend = ceil(static_cast<scalar_t>(h - roi_start_h + 1) / bin_size_h);
+			int pwstart = floor(static_cast<scalar_t>(w - roi_start_w) / bin_size_w);
+			int pwend = ceil(static_cast<scalar_t>(w - roi_start_w + 1) / bin_size_w);
+
+			phstart = min(max(phstart, 0), pooled_height);
+			phend = min(max(phend, 0), pooled_height);
+			pwstart = min(max(pwstart, 0), pooled_width);
+			pwend = min(max(pwend, 0), pooled_width);
+
+			for (int ph = phstart; ph < phend; ++ph) {
+				for (int pw = pwstart; pw < pwend; ++pw) {
+					if (offset_argmax_data[ph * pooled_width + pw] == (h * width + w)) {
+						gradient += offset_top_diff[ph * pooled_width + pw];
+					}
+				}
+			}
+		}
+		bottom_diff[index] = gradient;
+	}
+}
+
+int ROIRingPoolBackwardLaucher(const at::Tensor top_diff, const float spatial_scale, const int batch_size, const int num_rois,
+    const int height, const int width, const int channels, const int pooled_height,
+    const int pooled_width, const at::Tensor bottom_rois,
+    at::Tensor bottom_diff, const at::Tensor argmax_data)
+{
+    const int kThreadsPerBlock = 1024;
+    int output_size = batch_size * height * width * channels;
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+        top_diff.scalar_type(), "ROIRingPoolBackward", ([&] {
+            ROIRingPoolBackward<scalar_t><<<(output_size + kThreadsPerBlock - 1) / kThreadsPerBlock, kThreadsPerBlock>>>(
+                output_size, 
+                top_diff.data<scalar_t>(), 
+                argmax_data.data<int>(), 
+                num_rois,
+                scalar_t(spatial_scale), 
+                height, 
+                width, 
+                channels, 
+                pooled_height,
+                pooled_width, 
+                bottom_diff.data<scalar_t>(), 
+                bottom_rois.data<scalar_t>());
+        })
+    );
+    THCudaCheck(cudaGetLastError());
+    return 1;
+}
diff --git a/lib/roi_data_layer/__init__.py b/lib/roi_data_layer/__init__.py
new file mode 100644
index 0000000..7ba6a65
--- /dev/null
+++ b/lib/roi_data_layer/__init__.py
@@ -0,0 +1,6 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
diff --git a/lib/roi_data_layer/__pycache__/__init__.cpython-36.pyc b/lib/roi_data_layer/__pycache__/__init__.cpython-36.pyc
new file mode 100644
index 0000000..7943259
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/__init__.cpython-36.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/__init__.cpython-37.pyc b/lib/roi_data_layer/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..63bb7f9
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/layer.cpython-36.pyc b/lib/roi_data_layer/__pycache__/layer.cpython-36.pyc
new file mode 100644
index 0000000..caa45ae
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/layer.cpython-36.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/layer.cpython-37.pyc b/lib/roi_data_layer/__pycache__/layer.cpython-37.pyc
new file mode 100644
index 0000000..476ef50
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/layer.cpython-37.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/minibatch.cpython-36.pyc b/lib/roi_data_layer/__pycache__/minibatch.cpython-36.pyc
new file mode 100644
index 0000000..6c061d5
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/minibatch.cpython-36.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/minibatch.cpython-37.pyc b/lib/roi_data_layer/__pycache__/minibatch.cpython-37.pyc
new file mode 100644
index 0000000..930ea8a
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/minibatch.cpython-37.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/roidb.cpython-36.pyc b/lib/roi_data_layer/__pycache__/roidb.cpython-36.pyc
new file mode 100644
index 0000000..022ea94
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/roidb.cpython-36.pyc differ
diff --git a/lib/roi_data_layer/__pycache__/roidb.cpython-37.pyc b/lib/roi_data_layer/__pycache__/roidb.cpython-37.pyc
new file mode 100644
index 0000000..53bdbb7
Binary files /dev/null and b/lib/roi_data_layer/__pycache__/roidb.cpython-37.pyc differ
diff --git a/lib/roi_data_layer/augmentations.py b/lib/roi_data_layer/augmentations.py
new file mode 100644
index 0000000..efb03d0
--- /dev/null
+++ b/lib/roi_data_layer/augmentations.py
@@ -0,0 +1,409 @@
+import torch
+from torchvision import transforms
+import cv2
+import numpy as np
+import types
+from numpy import random
+from PIL import Image, ImageOps, ImageEnhance
+
+def intersect(box_a, box_b):
+    max_xy = np.minimum(box_a[:, 2:], box_b[2:])
+    min_xy = np.maximum(box_a[:, :2], box_b[:2])
+    inter = np.clip((max_xy - min_xy), a_min=0, a_max=np.inf)
+    return inter[:, 0] * inter[:, 1]
+
+
+def jaccard_numpy(box_a, box_b):
+    """Compute the jaccard overlap of two sets of boxes.  The jaccard overlap
+    is simply the intersection over union of two boxes.
+    E.g.:
+        A ∩ B / A ∪ B = A ∩ B / (area(A) + area(B) - A ∩ B)
+    Args:
+        box_a: Multiple bounding boxes, Shape: [num_boxes,4]
+        box_b: Single bounding box, Shape: [4]
+    Return:
+        jaccard overlap: Shape: [box_a.shape[0], box_a.shape[1]]
+    """
+    inter = intersect(box_a, box_b)
+    area_a = ((box_a[:, 2]-box_a[:, 0]) *
+              (box_a[:, 3]-box_a[:, 1]))  # [A,B]
+    area_b = ((box_b[2]-box_b[0]) *
+              (box_b[3]-box_b[1]))  # [A,B]
+    union = area_a + area_b - inter
+    return inter / union  # [A,B]
+
+
+class Compose(object):
+    """Composes several augmentations together.
+    Args:
+        transforms (List[Transform]): list of transforms to compose.
+    Example:
+        >>> augmentations.Compose([
+        >>>     transforms.CenterCrop(10),
+        >>>     transforms.ToTensor(),
+        >>> ])
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, img, boxes=None, labels=None):
+        for t in self.transforms:
+            img, boxes, labels = t(img, boxes, labels)
+        return img, boxes, labels
+
+
+class Lambda(object):
+    """Applies a lambda as a transform."""
+
+    def __init__(self, lambd):
+        assert isinstance(lambd, types.LambdaType)
+        self.lambd = lambd
+
+    def __call__(self, img, boxes=None, labels=None):
+        return self.lambd(img, boxes, labels)
+
+
+class ConvertFromInts(object):
+    def __call__(self, image, boxes=None, labels=None):
+        return image.astype(np.float32), boxes, labels
+
+
+class SubtractMeans(object):
+    def __init__(self, mean):
+        self.mean = np.array(mean, dtype=np.float32)
+
+    def __call__(self, image, boxes=None, labels=None):
+        image = image.astype(np.float32)
+        image -= self.mean
+        return image.astype(np.float32), boxes, labels
+
+
+class ToAbsoluteCoords(object):
+    def __call__(self, image, boxes=None, labels=None):
+        height, width, channels = image.shape
+        boxes[:, 0] *= width
+        boxes[:, 2] *= width
+        boxes[:, 1] *= height
+        boxes[:, 3] *= height
+
+        return image, boxes, labels
+
+
+class ToPercentCoords(object):
+    def __call__(self, image, boxes=None, labels=None):
+        height, width, channels = image.shape
+        boxes[:, 0] /= width
+        boxes[:, 2] /= width
+        boxes[:, 1] /= height
+        boxes[:, 3] /= height
+
+        return image, boxes, labels
+
+
+class Resize(object):
+    def __init__(self, size=300):
+        self.size = size
+
+    def __call__(self, image, boxes=None, labels=None):
+        image = cv2.resize(image, (self.size, self.size))
+        return image, boxes, labels
+
+
+class RandomSaturation(object):
+    def __init__(self, lower=0.5, upper=1.5):
+        self.lower = lower
+        self.upper = upper
+        assert self.upper >= self.lower, "contrast upper must be >= lower."
+        assert self.lower >= 0, "contrast lower must be non-negative."
+
+    def __call__(self, image, boxes=None, labels=None):
+        if random.randint(2):
+            image[:, :, 1] *= random.uniform(self.lower, self.upper)
+
+        return image, boxes, labels
+
+
+class RandomHue(object):
+    def __init__(self, delta=18.0):
+        assert delta >= 0.0 and delta <= 360.0
+        self.delta = delta
+
+    def __call__(self, image, boxes=None, labels=None):
+        if random.randint(2):
+            image[:, :, 0] += random.uniform(-self.delta, self.delta)
+            image[:, :, 0][image[:, :, 0] > 360.0] -= 360.0
+            image[:, :, 0][image[:, :, 0] < 0.0] += 360.0
+        return image, boxes, labels
+
+
+class RandomLightingNoise(object):
+    def __init__(self):
+        self.perms = ((0, 1, 2), (0, 2, 1),
+                      (1, 0, 2), (1, 2, 0),
+                      (2, 0, 1), (2, 1, 0))
+
+    def __call__(self, image, boxes=None, labels=None):
+        if random.randint(2):
+            swap = self.perms[random.randint(len(self.perms))]
+            shuffle = SwapChannels(swap)  # shuffle channels
+            image = shuffle(image)
+        return image, boxes, labels
+
+
+class ConvertColor(object):
+    def __init__(self, current='BGR', transform='HSV'):
+        self.transform = transform
+        self.current = current
+
+    def __call__(self, image, boxes=None, labels=None):
+        if self.current == 'BGR' and self.transform == 'HSV':
+            image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+        elif self.current == 'HSV' and self.transform == 'BGR':
+            image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
+        else:
+            raise NotImplementedError
+        return image, boxes, labels
+
+
+class RandomContrast(object):
+    def __init__(self, lower=0.5, upper=1.5):
+        self.lower = lower
+        self.upper = upper
+        assert self.upper >= self.lower, "contrast upper must be >= lower."
+        assert self.lower >= 0, "contrast lower must be non-negative."
+
+    # expects float image
+    def __call__(self, image, boxes=None, labels=None):
+        if random.randint(2):
+            alpha = random.uniform(self.lower, self.upper)
+            image *= alpha
+        return image, boxes, labels
+
+
+class RandomBrightness(object):
+    def __init__(self, delta=32):
+        assert delta >= 0.0
+        assert delta <= 255.0
+        self.delta = delta
+
+    def __call__(self, image, boxes=None, labels=None):
+        if random.randint(2):
+           delta = random.uniform(-self.delta, self.delta)
+           image += delta
+        return image, boxes, labels
+
+
+class ToCV2Image(object):
+    def __call__(self, tensor, boxes=None, labels=None):
+        return tensor.cpu().numpy().astype(np.float32).transpose((1, 2, 0)), boxes, labels
+
+
+class ToTensor(object):
+    def __call__(self, cvimage, boxes=None, labels=None):
+        return torch.from_numpy(cvimage.astype(np.float32)).permute(2, 0, 1), boxes, labels
+
+
+class RandomSampleCrop(object):
+    """Crop
+    Arguments:
+        img (Image): the image being input during training
+        boxes (Tensor): the original bounding boxes in pt form
+        labels (Tensor): the class labels for each bbox
+        mode (float tuple): the min and max jaccard overlaps
+    Return:
+        (img, boxes, classes)
+            img (Image): the cropped image
+            boxes (Tensor): the adjusted bounding boxes in pt form
+            labels (Tensor): the class labels for each bbox
+    """
+    def __init__(self):
+        self.sample_options = (
+            # using entire original input image
+            None,
+            # sample a patch s.t. MIN jaccard w/ obj in .1,.3,.4,.7,.9
+            (0.1, None),
+            (0.3, None),
+            (0.7, None),
+            (0.9, None),
+            # randomly sample a patch
+            (None, None),
+        )
+
+    def __call__(self, image, boxes=None, labels=None):
+        height, width, _ = image.shape
+        while True:
+            # randomly choose a mode
+            mode = random.choice(self.sample_options)
+            if mode is None:
+                return image, boxes, labels
+
+            min_iou, max_iou = mode
+            if min_iou is None:
+                min_iou = float('-inf')
+            if max_iou is None:
+                max_iou = float('inf')
+
+            # max trails (50)
+            for _ in range(50):
+                current_image = image
+
+                w = random.uniform(0.3 * width, width)
+                h = random.uniform(0.3 * height, height)
+
+                # aspect ratio constraint b/t .5 & 2
+                if h / w < 0.5 or h / w > 2:
+                    continue
+
+                left = random.uniform(width - w)
+                top = random.uniform(height - h)
+
+                # convert to integer rect x1,y1,x2,y2
+                rect = np.array([int(left), int(top), int(left+w), int(top+h)])
+
+                # calculate IoU (jaccard overlap) b/t the cropped and gt boxes
+                overlap = jaccard_numpy(boxes, rect)
+
+                # is min and max overlap constraint satisfied? if not try again
+                if overlap.min() < min_iou and max_iou < overlap.max():
+                    continue
+
+                # cut the crop from the image
+                current_image = current_image[rect[1]:rect[3], rect[0]:rect[2],
+                                              :]
+
+                # keep overlap with gt box IF center in sampled patch
+                centers = (boxes[:, :2] + boxes[:, 2:]) / 2.0
+
+                # mask in all gt boxes that above and to the left of centers
+                m1 = (rect[0] < centers[:, 0]) * (rect[1] < centers[:, 1])
+
+                # mask in all gt boxes that under and to the right of centers
+                m2 = (rect[2] > centers[:, 0]) * (rect[3] > centers[:, 1])
+
+                # mask in that both m1 and m2 are true
+                mask = m1 * m2
+
+                # have any valid boxes? try again if not
+                if not mask.any():
+                    continue
+
+                # take only matching gt boxes
+                current_boxes = boxes[mask, :].copy()
+
+                # take only matching gt labels
+                current_labels = labels[mask]
+
+                # should we use the box left and top corner or the crop's
+                current_boxes[:, :2] = np.maximum(current_boxes[:, :2],
+                                                  rect[:2])
+                # adjust to crop (by substracting crop's left,top)
+                current_boxes[:, :2] -= rect[:2]
+
+                current_boxes[:, 2:] = np.minimum(current_boxes[:, 2:],
+                                                  rect[2:])
+                # adjust to crop (by substracting crop's left,top)
+                current_boxes[:, 2:] -= rect[:2]
+
+                return current_image, current_boxes, current_labels
+
+
+class Expand(object):
+    def __init__(self, mean):
+        self.mean = mean
+
+    def __call__(self, image, boxes, labels):
+        if random.randint(2):
+            return image, boxes, labels
+
+        height, width, depth = image.shape
+        ratio = random.uniform(1, 4)
+        left = random.uniform(0, width*ratio - width)
+        top = random.uniform(0, height*ratio - height)
+
+        expand_image = np.zeros(
+            (int(height*ratio), int(width*ratio), depth),
+            dtype=image.dtype)
+        expand_image[:, :, :] = self.mean
+        expand_image[int(top):int(top + height),
+                     int(left):int(left + width)] = image
+        image = expand_image
+
+        boxes = boxes.copy()
+        boxes[:, :2] += (int(left), int(top))
+        boxes[:, 2:] += (int(left), int(top))
+
+        return image, boxes, labels
+
+
+class RandomMirror(object):
+    def __call__(self, image, boxes, classes):
+        _, width, _ = image.shape
+        if random.randint(2):
+            image = image[:, ::-1]
+            boxes = boxes.copy()
+            boxes[:, 0::2] = width - boxes[:, 2::-2]
+        return image, boxes, classes
+
+
+class SwapChannels(object):
+    """Transforms a tensorized image by swapping the channels in the order
+     specified in the swap tuple.
+    Args:
+        swaps (int triple): final order of channels
+            eg: (2, 1, 0)
+    """
+
+    def __init__(self, swaps):
+        self.swaps = swaps
+
+    def __call__(self, image):
+        """
+        Args:
+            image (Tensor): image tensor to be transformed
+        Return:
+            a tensor with channels swapped according to swap
+        """
+        # if torch.is_tensor(image):
+        #     image = image.data.cpu().numpy()
+        # else:
+        #     image = np.array(image)
+        image = image[:, :, self.swaps]
+        return image
+
+class PhotometricDistort(object):
+    def __init__(self):
+        self.pd = [
+            RandomContrast(),
+            ConvertColor(transform='HSV'),
+            RandomSaturation(),
+            RandomHue(),
+            ConvertColor(current='HSV', transform='BGR'),
+            RandomContrast()
+        ]
+        self.rand_brightness = RandomBrightness()
+        self.rand_light_noise = RandomLightingNoise()
+
+    def __call__(self, image, boxes, labels):
+        im = image.copy()
+        im, boxes, labels = self.rand_brightness(im, boxes, labels)
+        if random.randint(2):
+            distort = Compose(self.pd[:-1])
+        else:
+            distort = Compose(self.pd[1:])
+        im, boxes, labels = distort(im, boxes, labels)
+
+        return im, boxes, labels
+
+class SSDAugmentation(object):
+    def __init__(self, size=300, mean=(104, 117, 123)):
+        self.mean = mean
+        self.size = size
+        self.augment = Compose([
+            ConvertFromInts(),
+            PhotometricDistort()
+        ])
+
+    def __call__(self, img, boxes, labels):
+        return self.augment(img, boxes, labels)
+
diff --git a/lib/roi_data_layer/layer.py b/lib/roi_data_layer/layer.py
new file mode 100644
index 0000000..1ddc4c4
--- /dev/null
+++ b/lib/roi_data_layer/layer.py
@@ -0,0 +1,90 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick and Xinlei Chen
+# --------------------------------------------------------
+
+"""The data layer used during training to train a Fast R-CNN network.
+
+RoIDataLayer implements a Caffe Python layer.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from model.config import cfg
+from roi_data_layer.minibatch import get_minibatch
+from roi_data_layer.augmentations import SSDAugmentation
+import numpy as np
+import time
+
+class RoIDataLayer(object):
+  """Fast R-CNN data layer used for training."""
+
+  def __init__(self, roidb, num_classes, random=False):
+    """Set the roidb to be used by this layer during training."""
+    self._roidb = roidb
+    self._num_classes = num_classes
+    # Also set a random flag
+    self._random = random
+    self._shuffle_roidb_inds()
+    self._ssdaugmentation = SSDAugmentation()
+
+  def _shuffle_roidb_inds(self):
+    """Randomly permute the training roidb."""
+    # If the random flag is set, 
+    # then the database is shuffled according to system time
+    # Useful for the validation set
+    if self._random:
+      st0 = np.random.get_state()
+      millis = int(round(time.time() * 1000)) % 4294967295
+      np.random.seed(millis)
+    
+    if cfg.TRAIN.ASPECT_GROUPING:
+      widths = np.array([r['width'] for r in self._roidb])
+      heights = np.array([r['height'] for r in self._roidb])
+      horz = (widths >= heights)
+      vert = np.logical_not(horz)
+      horz_inds = np.where(horz)[0]
+      vert_inds = np.where(vert)[0]
+      inds = np.hstack((
+          np.random.permutation(horz_inds),
+          np.random.permutation(vert_inds)))
+      inds = np.reshape(inds, (-1, 2))
+      row_perm = np.random.permutation(np.arange(inds.shape[0]))
+      inds = np.reshape(inds[row_perm, :], (-1,))
+      self._perm = inds
+    else:
+      self._perm = np.random.permutation(np.arange(len(self._roidb)))
+    # Restore the random state
+    if self._random:
+      np.random.set_state(st0)
+      
+    self._cur = 0
+
+  def _get_next_minibatch_inds(self):
+    """Return the roidb indices for the next minibatch."""
+    
+    if self._cur + cfg.TRAIN.IMS_PER_BATCH >= len(self._roidb):
+      self._shuffle_roidb_inds()
+
+    db_inds = self._perm[self._cur:self._cur + cfg.TRAIN.IMS_PER_BATCH]
+    self._cur += cfg.TRAIN.IMS_PER_BATCH
+
+    return db_inds
+
+  def _get_next_minibatch(self, iter=None):
+    """Return the blobs to be used for the next minibatch.
+
+    If cfg.TRAIN.USE_PREFETCH is True, then blobs will be computed in a
+    separate process and made available through self._blob_queue.
+    """
+    db_inds = self._get_next_minibatch_inds()
+    minibatch_db = [self._roidb[i] for i in db_inds]
+    return get_minibatch(minibatch_db, self._num_classes, self._ssdaugmentation, iter)
+      
+  def forward(self, iter=None):
+    """Get blobs and copy them into this layer's top blob vector."""
+    blobs = self._get_next_minibatch(iter)
+    return blobs
diff --git a/lib/roi_data_layer/minibatch.py b/lib/roi_data_layer/minibatch.py
new file mode 100644
index 0000000..165d712
--- /dev/null
+++ b/lib/roi_data_layer/minibatch.py
@@ -0,0 +1,98 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick and Xinlei Chen
+# --------------------------------------------------------
+
+"""Compute minibatch blobs for training a Fast R-CNN network."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import numpy.random as npr
+import cv2
+from model.config import cfg
+from utils.blob import prep_im_for_blob, im_list_to_blob
+#import random
+from numpy import random
+
+def get_minibatch(roidb, num_classes, aug=None, iter=None):
+    """Given a roidb, construct a minibatch sampled from it."""
+    aug_time = 2
+
+    # Sample random scales to use for each image in this batch
+    blobs = {'data': [], 'gt_boxes': [], 'im_info': [], 'image_level_labels': None, 'ss_boxes': [], 'flag': None}
+
+    zero_index = np.arange(len(cfg.TRAIN.SCALES))
+    dumy_index = npr.choice(zero_index, size=aug_time, replace=False)
+
+    for i in range(aug_time):
+        random_scale_inds = dumy_index[i]
+        target_size = cfg.TRAIN.SCALES[random_scale_inds]
+
+        # Get the input image blob, formatted for caffe
+        im_blob, im_scales = _get_image_blob(roidb, random_scale_inds, aug, target_size, iter)
+
+        blobs['data'].append(im_blob)
+
+        assert len(im_scales) == 1, "Single batch only"
+        assert len(roidb) == 1, "Single batch only"
+
+        blobs['im_info'].append(np.array([im_blob.shape[1], im_blob.shape[2], im_scales[0]], dtype=np.float32))
+        blobs['image_level_labels'] = roidb[0]['image_level_labels']
+
+        # add ss_boxes into blob
+        # Changed for WSDNN
+        ss_inds = np.where(roidb[0]['gt_classes'] == -1)[0]  # remove gt_rois in ss_boxes
+        ss_boxes = np.empty((len(ss_inds), 5), dtype=np.float32)
+        ss_boxes[:, 1:] = roidb[0]['boxes'][ss_inds, :] * im_scales[0]
+        ss_boxes[:, 0] = 0
+        blobs['gt_boxes'].append(ss_boxes.copy())
+
+        blobs['ss_boxes'].append(ss_boxes)
+        ss_boxes_flip = ss_boxes.copy()
+        width = im_blob.shape[2]
+        num_boxes_each = ss_boxes.shape[0]
+        oldx1 = ss_boxes_flip[:, 1].copy()
+        oldx2 = ss_boxes_flip[:, 3].copy()
+        width = np.ones((num_boxes_each)) * width
+        ss_boxes_flip[:, 1] = width - oldx2
+        ss_boxes_flip[:, 3] = width - oldx1
+        blobs['ss_boxes'].append(ss_boxes_flip)
+
+    return blobs
+
+
+def _get_image_blob(roidb, scale_inds, aug, target_size, iter=None):
+    """Builds an input blob from the images in the roidb at the specified
+  scales.
+  """
+    num_images = len(roidb)
+    processed_ims = []
+    im_scales = []
+    for i in range(num_images):
+        im = cv2.imread(roidb[i]['image'])
+        if iter >= (cfg.TRAIN.STEP_ITERS - cfg.TRAIN.SNAPSHOT_ITERS):
+            im, _, _ = aug(im, None, None)
+
+        if roidb[i]['flipped']:
+            im_flip = im
+            im = im[:, ::-1, :]
+        else:
+            im_flip = im[:, ::-1, :]
+
+        im, im_scale = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size,
+                                        cfg.TRAIN.MAX_SIZE)
+        im_scales.append(im_scale)
+
+        im_flip, _ = prep_im_for_blob(im_flip, cfg.PIXEL_MEANS, target_size,
+                                      cfg.TRAIN.MAX_SIZE)
+        processed_ims.append(im)
+        processed_ims.append(im_flip)
+
+    # Create a blob to hold the input images
+    blob = im_list_to_blob(processed_ims)
+
+    return blob, im_scales
diff --git a/lib/roi_data_layer/roidb.py b/lib/roi_data_layer/roidb.py
new file mode 100644
index 0000000..42c6a24
--- /dev/null
+++ b/lib/roi_data_layer/roidb.py
@@ -0,0 +1,47 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+"""Transform a roidb into a trainable roidb by adding a bunch of metadata."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from model.config import cfg
+import PIL
+
+def prepare_roidb(imdb):
+  """Enrich the imdb's roidb by adding some derived quantities that
+  are useful for training. This function precomputes the maximum
+  overlap, taken over ground-truth boxes, between each ROI and
+  each ground-truth box. The class with maximum overlap is also
+  recorded.
+  """
+  roidb = imdb.roidb
+  if not (imdb.name.startswith('coco')):
+    sizes = [PIL.Image.open(imdb.image_path_at(i)).size
+         for i in range(imdb.num_images)]
+  for i in range(len(imdb.image_index)):
+    roidb[i]['image'] = imdb.image_path_at(i)
+    if not (imdb.name.startswith('coco')):
+      roidb[i]['width'] = sizes[i][0]
+      roidb[i]['height'] = sizes[i][1]
+    # need gt_overlaps as a dense array for argmax
+    gt_overlaps = roidb[i]['gt_overlaps'].toarray()
+    # max overlap with gt over classes (columns)
+    max_overlaps = gt_overlaps.max(axis=1)
+    # gt class that had the max overlap
+    max_classes = gt_overlaps.argmax(axis=1)
+    roidb[i]['max_classes'] = max_classes
+    roidb[i]['max_overlaps'] = max_overlaps
+    # sanity checks
+    # max overlap of 0 => class should be zero (background)
+    zero_inds = np.where(max_overlaps == 0)[0]
+    assert all(max_classes[zero_inds] == 0)
+    # max overlap > 0 => class should not be zero (must be a fg class)
+    nonzero_inds = np.where(max_overlaps > 0)[0]
+    #assert all(max_classes[nonzero_inds] != 0)
diff --git a/lib/setup_cuda.py b/lib/setup_cuda.py
new file mode 100644
index 0000000..e9ec2cc
--- /dev/null
+++ b/lib/setup_cuda.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 29 11:27:44 2019
+
+@author: vasgaoweithu
+"""
+
+import os
+import platform
+from setuptools import Extension, dist, find_packages, setup
+
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+dist.Distribution().fetch_build_eggs(['Cython', 'numpy>=1.11.1'])
+import numpy as np
+from Cython.Build import cythonize
+
+
+def make_cuda_ext(name, module, sources):
+
+    return CUDAExtension(
+        name='{}.{}'.format(module, name),
+        sources=[os.path.join(*module.split('.'), p) for p in sources],
+        extra_compile_args={
+            'cxx': [],
+            'nvcc': [
+                '-D__CUDA_NO_HALF_OPERATORS__',
+                '-D__CUDA_NO_HALF_CONVERSIONS__',
+                '-D__CUDA_NO_HALF2_OPERATORS__',
+            ]
+        })
+
+    
+def make_cython_ext(name, module, sources):
+    extra_compile_args = None
+    if platform.system() != 'Windows':
+        extra_compile_args = {
+            'cxx': ['-Wno-unused-function', '-Wno-write-strings']
+        }
+
+    extension = Extension(
+        '{}.{}'.format(module, name),
+        [os.path.join(*module.split('.'), p) for p in sources],
+        include_dirs=[np.get_include()],
+        language='c++',
+        extra_compile_args=extra_compile_args)
+    extension, = cythonize(extension)
+    return extension
+
+if __name__ == '__main__':
+    setup(
+        name="faster rcnn",
+        version="0.1",
+        author="GaoWei",
+        description="Weakly Supervised Obejct Detection",
+        packages=find_packages(exclude=("cfgs", "tools")),
+        ext_modules=[
+# =============================================================================
+#             make_cython_ext(
+#                 name='soft_nms_cpu',
+#                 module='ops.nms',
+#                 sources=['src/soft_nms_cpu.pyx']),
+#             make_cuda_ext(
+#                 name='nms_cpu',
+#                 module='ops.nms',
+#                 sources=['src/nms_cpu.cpp']),
+#             make_cuda_ext(
+#                 name='nms_cuda',
+#                 module='ops.nms',
+#                 sources=['src/nms_cuda.cpp', 'src/nms_kernel.cu']),
+#             make_cuda_ext(
+#                 name='roi_align_cuda',
+#                 module='ops.roi_align',
+#                 sources=['src/roi_align_cuda.cpp', 'src/roi_align_kernel.cu']),
+#             make_cuda_ext(
+#                 name='roi_pool_cuda',
+#                 module='ops.roi_pool',
+#                 sources=['src/roi_pool_cuda.cpp', 'src/roi_pool_kernel.cu']),
+#             make_cuda_ext(
+#                     name='roi_crop_cpu',
+#                     module='ops.roi_crop',
+#                     sources=['src/roi_crop_cpu.cpp']
+#                     ),
+#             make_cuda_ext(
+#                     name='roi_crop_cuda',
+#                     module='ops.roi_crop',
+#                     sources=['src/roi_crop_cuda.cpp', 'src/roi_crop_kernel.cu']
+#                     ),
+# =============================================================================
+            make_cuda_ext(
+                    name='roi_ring_pool_cuda',
+                    module='ops.roi_ring_pool',
+                    sources=['src/roi_ring_pool_cuda.cpp', 'src/roi_ring_pool_kernel.cu']
+                    )
+                    ],
+            cmdclass={'build_ext': BuildExtension},
+            zip_safe=False
+            )
\ No newline at end of file
diff --git a/lib/utils/.gitignore b/lib/utils/.gitignore
new file mode 100644
index 0000000..3347b69
--- /dev/null
+++ b/lib/utils/.gitignore
@@ -0,0 +1,4 @@
+*.c
+*.cpp
+*.h
+*.hpp
diff --git a/lib/utils/__init__.py b/lib/utils/__init__.py
new file mode 100644
index 0000000..7ba6a65
--- /dev/null
+++ b/lib/utils/__init__.py
@@ -0,0 +1,6 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
diff --git a/lib/utils/__pycache__/__init__.cpython-36.pyc b/lib/utils/__pycache__/__init__.cpython-36.pyc
new file mode 100644
index 0000000..d41797f
Binary files /dev/null and b/lib/utils/__pycache__/__init__.cpython-36.pyc differ
diff --git a/lib/utils/__pycache__/__init__.cpython-37.pyc b/lib/utils/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000..b0b8251
Binary files /dev/null and b/lib/utils/__pycache__/__init__.cpython-37.pyc differ
diff --git a/lib/utils/__pycache__/bbox.cpython-36.pyc b/lib/utils/__pycache__/bbox.cpython-36.pyc
new file mode 100644
index 0000000..304ae68
Binary files /dev/null and b/lib/utils/__pycache__/bbox.cpython-36.pyc differ
diff --git a/lib/utils/__pycache__/bbox.cpython-37.pyc b/lib/utils/__pycache__/bbox.cpython-37.pyc
new file mode 100644
index 0000000..1c87d63
Binary files /dev/null and b/lib/utils/__pycache__/bbox.cpython-37.pyc differ
diff --git a/lib/utils/__pycache__/blob.cpython-36.pyc b/lib/utils/__pycache__/blob.cpython-36.pyc
new file mode 100644
index 0000000..6675518
Binary files /dev/null and b/lib/utils/__pycache__/blob.cpython-36.pyc differ
diff --git a/lib/utils/__pycache__/blob.cpython-37.pyc b/lib/utils/__pycache__/blob.cpython-37.pyc
new file mode 100644
index 0000000..c9ebfb6
Binary files /dev/null and b/lib/utils/__pycache__/blob.cpython-37.pyc differ
diff --git a/lib/utils/__pycache__/timer.cpython-36.pyc b/lib/utils/__pycache__/timer.cpython-36.pyc
new file mode 100644
index 0000000..1f4fff8
Binary files /dev/null and b/lib/utils/__pycache__/timer.cpython-36.pyc differ
diff --git a/lib/utils/__pycache__/timer.cpython-37.pyc b/lib/utils/__pycache__/timer.cpython-37.pyc
new file mode 100644
index 0000000..37f5652
Binary files /dev/null and b/lib/utils/__pycache__/timer.cpython-37.pyc differ
diff --git a/lib/utils/__pycache__/visualization.cpython-36.pyc b/lib/utils/__pycache__/visualization.cpython-36.pyc
new file mode 100644
index 0000000..7897a65
Binary files /dev/null and b/lib/utils/__pycache__/visualization.cpython-36.pyc differ
diff --git a/lib/utils/__pycache__/visualization.cpython-37.pyc b/lib/utils/__pycache__/visualization.cpython-37.pyc
new file mode 100644
index 0000000..06be3e2
Binary files /dev/null and b/lib/utils/__pycache__/visualization.cpython-37.pyc differ
diff --git a/lib/utils/bbox.py b/lib/utils/bbox.py
new file mode 100644
index 0000000..676d374
--- /dev/null
+++ b/lib/utils/bbox.py
@@ -0,0 +1,30 @@
+import torch
+import numpy as np
+
+def bbox_overlaps(boxes, query_boxes):
+    """
+    Parameters
+    ----------
+    boxes: (N, 4) ndarray or tensor or variable
+    query_boxes: (K, 4) ndarray or tensor or variable
+    Returns
+    -------
+    overlaps: (N, K) overlap between boxes and query_boxes
+    """
+    if isinstance(boxes, np.ndarray):
+        boxes = torch.from_numpy(boxes)
+        query_boxes = torch.from_numpy(query_boxes)
+        out_fn = lambda x: x.numpy() # If input is ndarray, turn the overlaps back to ndarray when return
+    else:
+        out_fn = lambda x: x
+
+    box_areas = (boxes[:, 2] - boxes[:, 0] + 1) * \
+            (boxes[:, 3] - boxes[:, 1] + 1)
+    query_areas = (query_boxes[:, 2] - query_boxes[:, 0] + 1) * \
+            (query_boxes[:, 3] - query_boxes[:, 1] + 1)
+
+    iw = (torch.min(boxes[:, 2:3], query_boxes[:, 2:3].t()) - torch.max(boxes[:, 0:1], query_boxes[:, 0:1].t()) + 1).clamp(min=0)
+    ih = (torch.min(boxes[:, 3:4], query_boxes[:, 3:4].t()) - torch.max(boxes[:, 1:2], query_boxes[:, 1:2].t()) + 1).clamp(min=0)
+    ua = box_areas.view(-1, 1) + query_areas.view(1, -1) - iw * ih
+    overlaps = iw * ih / ua
+    return out_fn(overlaps)
\ No newline at end of file
diff --git a/lib/utils/blob.py b/lib/utils/blob.py
new file mode 100644
index 0000000..0399fdd
--- /dev/null
+++ b/lib/utils/blob.py
@@ -0,0 +1,47 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+"""Blob helper functions."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import cv2
+
+
+def im_list_to_blob(ims):
+  """Convert a list of images into a network input.
+
+  Assumes images are already prepared (means subtracted, BGR order, ...).
+  """
+  max_shape = np.array([im.shape for im in ims]).max(axis=0)
+  num_images = len(ims)
+  blob = np.zeros((num_images, max_shape[0], max_shape[1], 3),
+                  dtype=np.float32)
+  for i in range(num_images):
+    im = ims[i]
+    blob[i, 0:im.shape[0], 0:im.shape[1], :] = im
+
+  return blob
+
+
+def prep_im_for_blob(im, pixel_means, target_size, max_size):
+  """Mean subtract and scale an image for use in a blob."""
+  im = im.astype(np.float32, copy=False)
+  im -= pixel_means
+  im_shape = im.shape
+  im_size_min = np.min(im_shape[0:2])
+  im_size_max = np.max(im_shape[0:2])
+  im_scale = float(target_size) / float(im_size_min)
+  # Prevent the biggest axis from being more than MAX_SIZE
+  if np.round(im_scale * im_size_max) > max_size:
+    im_scale = float(max_size) / float(im_size_max)
+  im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale,
+                  interpolation=cv2.INTER_LINEAR)
+
+  return im, im_scale
diff --git a/lib/utils/timer.py b/lib/utils/timer.py
new file mode 100644
index 0000000..b940754
--- /dev/null
+++ b/lib/utils/timer.py
@@ -0,0 +1,45 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+import time
+import torch
+
+class Timer(object):
+    """A simple timer."""
+    def __init__(self):
+        self._total_time = {}
+        self._calls = {}
+        self._start_time = {}
+        self._diff = {}
+        self._average_time = {}
+
+    def tic(self, name='default'):
+        # using time.time instead of time.clock because time time.clock
+        # does not normalize for multithreading
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        self._start_time[name] = time.time()
+
+    def toc(self, name='default', average=True):
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        self._diff[name] = time.time() - self._start_time[name]
+        self._total_time[name] = self._total_time.get(name, 0.) + self._diff[name]
+        self._calls[name] = self._calls.get(name, 0 ) + 1
+        self._average_time[name] = self._total_time[name] / self._calls[name]
+        if average:
+            return self._average_time[name]
+        else:
+            return self._diff[name]
+
+    def average_time(self, name='default'):
+        return self._average_time[name]
+
+    def total_time(self, name='default'):
+        return self._total_time[name]
+
+timer = Timer()
diff --git a/lib/utils/visualization.py b/lib/utils/visualization.py
new file mode 100644
index 0000000..ed4b0fb
--- /dev/null
+++ b/lib/utils/visualization.py
@@ -0,0 +1,89 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from six.moves import range
+import PIL.Image as Image
+import PIL.ImageColor as ImageColor
+import PIL.ImageDraw as ImageDraw
+import PIL.ImageFont as ImageFont
+
+STANDARD_COLORS = [
+    'AliceBlue', 'Chartreuse', 'Aqua', 'Aquamarine', 'Azure', 'Beige', 'Bisque',
+    'BlanchedAlmond', 'BlueViolet', 'BurlyWood', 'CadetBlue', 'AntiqueWhite',
+    'Chocolate', 'Coral', 'CornflowerBlue', 'Cornsilk', 'Crimson', 'Cyan',
+    'DarkCyan', 'DarkGoldenRod', 'DarkGrey', 'DarkKhaki', 'DarkOrange',
+    'DarkOrchid', 'DarkSalmon', 'DarkSeaGreen', 'DarkTurquoise', 'DarkViolet',
+    'DeepPink', 'DeepSkyBlue', 'DodgerBlue', 'FireBrick', 'FloralWhite',
+    'ForestGreen', 'Fuchsia', 'Gainsboro', 'GhostWhite', 'Gold', 'GoldenRod',
+    'Salmon', 'Tan', 'HoneyDew', 'HotPink', 'IndianRed', 'Ivory', 'Khaki',
+    'Lavender', 'LavenderBlush', 'LawnGreen', 'LemonChiffon', 'LightBlue',
+    'LightCoral', 'LightCyan', 'LightGoldenRodYellow', 'LightGray', 'LightGrey',
+    'LightGreen', 'LightPink', 'LightSalmon', 'LightSeaGreen', 'LightSkyBlue',
+    'LightSlateGray', 'LightSlateGrey', 'LightSteelBlue', 'LightYellow', 'Lime',
+    'LimeGreen', 'Linen', 'Magenta', 'MediumAquaMarine', 'MediumOrchid',
+    'MediumPurple', 'MediumSeaGreen', 'MediumSlateBlue', 'MediumSpringGreen',
+    'MediumTurquoise', 'MediumVioletRed', 'MintCream', 'MistyRose', 'Moccasin',
+    'NavajoWhite', 'OldLace', 'Olive', 'OliveDrab', 'Orange', 'OrangeRed',
+    'Orchid', 'PaleGoldenRod', 'PaleGreen', 'PaleTurquoise', 'PaleVioletRed',
+    'PapayaWhip', 'PeachPuff', 'Peru', 'Pink', 'Plum', 'PowderBlue', 'Purple',
+    'Red', 'RosyBrown', 'RoyalBlue', 'SaddleBrown', 'Green', 'SandyBrown',
+    'SeaGreen', 'SeaShell', 'Sienna', 'Silver', 'SkyBlue', 'SlateBlue',
+    'SlateGray', 'SlateGrey', 'Snow', 'SpringGreen', 'SteelBlue', 'GreenYellow',
+    'Teal', 'Thistle', 'Tomato', 'Turquoise', 'Violet', 'Wheat', 'White',
+    'WhiteSmoke', 'Yellow', 'YellowGreen'
+]
+
+NUM_COLORS = len(STANDARD_COLORS)
+
+try:
+  FONT = ImageFont.truetype('arial.ttf', 24)
+except IOError:
+  FONT = ImageFont.load_default()
+
+def _draw_single_box(image, xmin, ymin, xmax, ymax, display_str, font, color='black', thickness=4):
+  draw = ImageDraw.Draw(image)
+  (left, right, top, bottom) = (xmin, xmax, ymin, ymax)
+  draw.line([(left, top), (left, bottom), (right, bottom),
+             (right, top), (left, top)], width=thickness, fill=color)
+  text_bottom = bottom
+  # Reverse list and print from bottom to top.
+  text_width, text_height = font.getsize(display_str)
+  margin = np.ceil(0.05 * text_height)
+  draw.rectangle(
+      [(left, text_bottom - text_height - 2 * margin), (left + text_width,
+                                                        text_bottom)],
+      fill=color)
+  draw.text(
+      (left + margin, text_bottom - text_height - margin),
+      display_str,
+      fill='black',
+      font=font)
+
+  return image
+
+def draw_bounding_boxes(image, gt_boxes, im_info):
+  num_boxes = gt_boxes.shape[0]
+  gt_boxes_new = gt_boxes.copy()
+  gt_boxes_new[:,:4] = np.round(gt_boxes_new[:,:4].copy() / im_info[2])
+  disp_image = Image.fromarray(np.uint8(image[0]))
+
+  for i in range(num_boxes):
+    this_class = int(gt_boxes_new[i, 4])
+    disp_image = _draw_single_box(disp_image, 
+                                gt_boxes_new[i, 0],
+                                gt_boxes_new[i, 1],
+                                gt_boxes_new[i, 2],
+                                gt_boxes_new[i, 3],
+                                'N%02d-C%02d' % (i, this_class),
+                                FONT,
+                                color=STANDARD_COLORS[this_class % NUM_COLORS])
+
+  image[0, :] = np.array(disp_image)
+  return image
\ No newline at end of file
diff --git a/output/train_faster_rcnn.sh b/output/train_faster_rcnn.sh
new file mode 100644
index 0000000..ca1a081
--- /dev/null
+++ b/output/train_faster_rcnn.sh
@@ -0,0 +1,88 @@
+#!/bin/bash
+
+alias time='/usr/bin/time'
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[50000]"
+    ITERS=100000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[80000]"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    STEPSIZE="[350000]"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}_${NET}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_MELM_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/default/${NET}_MELM_iter_${ITERS}.pth
+fi
+set -x
+
+if [ ! -f ${NET_FINAL}.index ]; then
+  if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --tag ${EXTRA_ARGS_SLUG} \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  else
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  fi
+fi
+
+echo $@
+./experiments/scripts/test_faster_rcnn.sh $@
diff --git a/tensorboard/train_faster_rcnn.sh b/tensorboard/train_faster_rcnn.sh
new file mode 100644
index 0000000..ca1a081
--- /dev/null
+++ b/tensorboard/train_faster_rcnn.sh
@@ -0,0 +1,88 @@
+#!/bin/bash
+
+alias time='/usr/bin/time'
+
+set -x
+set -e
+
+export PYTHONUNBUFFERED="True"
+
+GPU_ID=$1
+DATASET=$2
+NET=$3
+
+array=( $@ )
+len=${#array[@]}
+EXTRA_ARGS=${array[@]:3:$len}
+EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
+
+case ${DATASET} in
+  pascal_voc)
+    TRAIN_IMDB="voc_2007_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[50000]"
+    ITERS=100000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  pascal_voc_0712)
+    TRAIN_IMDB="voc_2007_trainval+voc_2012_trainval"
+    TEST_IMDB="voc_2007_test"
+    STEPSIZE="[80000]"
+    ITERS=110000
+    ANCHORS="[8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  coco)
+    TRAIN_IMDB="coco_2014_train+coco_2014_valminusminival"
+    TEST_IMDB="coco_2014_minival"
+    STEPSIZE="[350000]"
+    ITERS=490000
+    ANCHORS="[4,8,16,32]"
+    RATIOS="[0.5,1,2]"
+    ;;
+  *)
+    echo "No dataset given"
+    exit
+    ;;
+esac
+
+LOG="experiments/logs/${NET}_${TRAIN_IMDB}_${EXTRA_ARGS_SLUG}_${NET}.txt.`date +'%Y-%m-%d_%H-%M-%S'`"
+exec &> >(tee -a "$LOG")
+echo Logging output to "$LOG"
+
+set +x
+if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/${EXTRA_ARGS_SLUG}/${NET}_MELM_iter_${ITERS}.pth
+else
+  NET_FINAL=output/${NET}/${TRAIN_IMDB}/default/${NET}_MELM_iter_${ITERS}.pth
+fi
+set -x
+
+if [ ! -f ${NET_FINAL}.index ]; then
+  if [[ ! -z  ${EXTRA_ARGS_SLUG}  ]]; then
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --tag ${EXTRA_ARGS_SLUG} \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  else
+    CUDA_VISIBLE_DEVICES=${GPU_ID} python ./tools/trainval_net.py \
+      --weight data/imagenet_weights/${NET}.pth \
+      --imdb ${TRAIN_IMDB} \
+      --imdbval ${TEST_IMDB} \
+      --iters ${ITERS} \
+      --cfg experiments/cfgs/${NET}.yml \
+      --net ${NET} \
+      --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} \
+      TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}
+  fi
+fi
+
+echo $@
+./experiments/scripts/test_faster_rcnn.sh $@
diff --git a/tools/__pycache__/_init_paths.cpython-36.pyc b/tools/__pycache__/_init_paths.cpython-36.pyc
new file mode 100644
index 0000000..364ef85
Binary files /dev/null and b/tools/__pycache__/_init_paths.cpython-36.pyc differ
diff --git a/tools/__pycache__/_init_paths.cpython-37.pyc b/tools/__pycache__/_init_paths.cpython-37.pyc
new file mode 100644
index 0000000..aa6e5ee
Binary files /dev/null and b/tools/__pycache__/_init_paths.cpython-37.pyc differ
diff --git a/tools/_init_paths.py b/tools/_init_paths.py
new file mode 100644
index 0000000..99550c5
--- /dev/null
+++ b/tools/_init_paths.py
@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Nov 16 15:17:33 2018
+
+@author: vasgaoweithu
+"""
+
+import os.path as osp
+import sys
+def add_path(path):
+    if path not in sys.path:
+        sys.path.insert(0, path)
+this_dir = osp.dirname(__file__)
+
+lib_path = osp.join(this_dir, '..', 'lib')
+add_path(lib_path)
+
+coco_path = osp.join(this_dir, '..', 'data', 'coco', 'PythonAPI')
+add_path(coco_path)
diff --git a/tools/demo.py b/tools/demo.py
new file mode 100644
index 0000000..634106e
--- /dev/null
+++ b/tools/demo.py
@@ -0,0 +1,182 @@
+#!/usr/bin/env python
+
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Xinlei Chen, based on code from Ross Girshick
+# --------------------------------------------------------
+
+"""
+Demo script showing detections in sample images.
+
+See README.md for installation instructions before running.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import _init_paths
+from model.config import cfg
+from model.test import im_detect
+from model.nms_wrapper import nms
+from datasets.factory import get_imdb
+
+from utils.timer import Timer
+import matplotlib.pyplot as plt
+import numpy as np
+import os, cv2
+import argparse
+
+from nets.vgg16 import vgg16
+from nets.resnet_v1 import resnetv1
+
+import torch
+
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+CLASSES = ('__background__',
+           'aeroplane', 'bicycle', 'bird', 'boat',
+           'bottle', 'bus', 'car', 'cat', 'chair',
+           'cow', 'diningtable', 'dog', 'horse',
+           'motorbike', 'person', 'pottedplant',
+           'sheep', 'sofa', 'train', 'tvmonitor')
+CLASSES = (
+    'aeroplane', 'bicycle', 'bird', 'boat',
+    'bottle', 'bus', 'car', 'cat', 'chair',
+    'cow', 'diningtable', 'dog', 'horse',
+    'motorbike', 'person', 'pottedplant',
+    'sheep', 'sofa', 'train', 'tvmonitor')
+
+NETS = {'vgg16': ('vgg16_faster_rcnn_iter_%d.pth',), 'res101': ('res101_faster_rcnn_iter_%d.pth',)}
+DATASETS = {'pascal_voc': ('voc_2007_trainval',), 'pascal_voc_0712': ('voc_2007_trainval+voc_2012_trainval',)}
+
+
+def vis_detections(im, class_name, dets, thresh=0.5):
+    """Draw detected bounding boxes."""
+    inds = np.where(dets[:, -1] >= thresh)[0]
+    if len(inds) == 0:
+        # print('hahaha')
+        return
+
+    im = im[:, :, (2, 1, 0)]
+    fig, ax = plt.subplots(figsize=(12, 12))
+    ax.imshow(im, aspect='equal')
+    for i in inds:
+        bbox = dets[i, :4]
+        score = dets[i, -1]
+
+        ax.add_patch(
+            plt.Rectangle((bbox[0], bbox[1]),
+                          bbox[2] - bbox[0],
+                          bbox[3] - bbox[1], fill=False,
+                          edgecolor='red', linewidth=3.5)
+        )
+        ax.text(bbox[0], bbox[1] - 2,
+                '{:s} {:.3f}'.format(class_name, score),
+                bbox=dict(facecolor='blue', alpha=0.5),
+                fontsize=14, color='white')
+
+    ax.set_title(('{} detections with '
+                  'p({} | box) >= {:.1f}').format(class_name, class_name,
+                                                  thresh),
+                 fontsize=14)
+    plt.axis('off')
+    plt.tight_layout()
+    plt.draw()
+
+
+def demo(net, image_name, roidb):
+    """Detect object classes in an image using pre-computed object proposals."""
+
+    # Load the demo image
+    # im_file = os.path.join(cfg.DATA_DIR, 'demo', image_name)
+    im = cv2.imread(image_name)
+
+    # Detect all object classes and regress object bounds
+    timer = Timer()
+    timer.tic()
+    scores, boxes, _, _ = im_detect(net, im, roidb)
+    timer.toc()
+    print('Detection took {:.3f}s for {:d} object proposals'.format(timer.total_time(), boxes.shape[0]))
+
+    # Visualize detections for each class
+    CONF_THRESH = 0.3
+    NMS_THRESH = 0.3
+    for cls_ind, cls in enumerate(CLASSES):
+        # cls_ind += 1 # because we skipped background
+        cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
+        cls_scores = scores[:, cls_ind]
+        dets = np.hstack((cls_boxes,
+                          cls_scores[:, np.newaxis])).astype(np.float32)
+        keep = nms(torch.from_numpy(dets), NMS_THRESH)
+        dets = dets[keep.numpy(), :]
+        vis_detections(im, cls, dets, thresh=CONF_THRESH)
+
+
+def parse_args():
+    """Parse input arguments."""
+    parser = argparse.ArgumentParser(description='Tensorflow Faster R-CNN demo')
+    parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16 res101]',
+                        choices=NETS.keys(), default='res101')
+    parser.add_argument('--dataset', dest='dataset', help='Trained dataset [pascal_voc pascal_voc_0712]',
+                        choices=DATASETS.keys(), default='pascal_voc_0712')
+    args = parser.parse_args()
+
+    return args
+
+
+if __name__ == '__main__':
+    cfg.TEST.HAS_RPN = True  # Use RPN for proposals
+    args = parse_args()
+
+    # model path
+    demonet = args.demo_net
+    dataset = args.dataset
+    saved_model = os.path.join('output', demonet, DATASETS[dataset][0], 'default',
+                               NETS[demonet][0] % (100000 if dataset == 'pascal_voc' else 110000))
+
+    if not os.path.isfile(saved_model):
+        raise IOError(('{:s} not found.\nDid you download the proper networks from '
+                       'our server and place them properly?').format(saved_model))
+
+    if args.dataset == 'pascal_voc':
+        test_name = 'voc_2007_test'
+
+    imdb = get_imdb(test_name)
+    roidb = imdb.roidb
+    # imdb.competition_mode(args.comp_mode)
+    # load network
+    if demonet == 'vgg16':
+        net = vgg16()
+    elif demonet == 'res101':
+        net = resnetv1(num_layers=101)
+    else:
+        raise NotImplementedError
+    net.create_architecture(20,
+                            tag='default', anchor_scales=[8, 16, 32])
+
+    net.load_state_dict(torch.load(saved_model, map_location=lambda storage, loc: storage))
+
+    net.eval()
+    if not torch.cuda.is_available():
+        net._device = 'cpu'
+    net.to(net._device)
+
+    print('Loaded network {:s}'.format(saved_model))
+
+    # im_names = ['000456.jpg', '000542.jpg', '001150.jpg',
+    #            '001763.jpg', '004545.jpg']
+    # for im_name in im_names:
+    #    print('~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~')
+    #    print('Demo for data/demo/{}'.format(im_name))
+    #    demo(net, im_name)
+
+    index = np.random.randint(0, len(imdb.image_index), 10)
+
+    index = np.arange(10, 20)
+
+    for i in index:
+        im_path = imdb.image_path_at(i)
+        demo(net, im_path, roidb[i])
+
+    plt.show()
diff --git a/tools/test_net.py b/tools/test_net.py
new file mode 100644
index 0000000..cfa4d47
--- /dev/null
+++ b/tools/test_net.py
@@ -0,0 +1,146 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Zheqi he, Xinlei Chen, based on code from Ross Girshick
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import _init_paths
+from model.test import test_net
+from model.train_val import get_training_roidb
+from model.config import cfg, cfg_from_file, cfg_from_list
+from datasets.factory import get_imdb
+import argparse
+import pprint
+import time, os, sys
+
+from nets.vgg16 import MELM_vgg16
+from nets.resnet_v1 import resnetv1
+from nets.mobilenet_v1 import mobilenetv1
+
+import torch
+
+def parse_args():
+  """
+  Parse input arguments
+  """
+  parser = argparse.ArgumentParser(description='Test a Fast R-CNN network')
+  parser.add_argument('--cfg', dest='cfg_file',
+            help='optional config file', default=None, type=str)
+  parser.add_argument('--model', dest='model',
+            help='model to test',
+            default=None, type=str)
+  parser.add_argument('--imdb', dest='imdb_name',
+            help='dataset to test',
+            default='voc_2007_test', type=str)
+  parser.add_argument('--comp', dest='comp_mode', help='competition mode',
+            action='store_true')
+  parser.add_argument('--num_dets', dest='max_per_image',
+            help='max number of detections per image',
+            default=100, type=int)
+  parser.add_argument('--tag', dest='tag',
+                        help='tag of the model',
+                        default='', type=str)
+  parser.add_argument('--net', dest='net',
+                      help='vgg16, res50, res101, res152, mobile',
+                      default='res50', type=str)
+  parser.add_argument('--set', dest='set_cfgs',
+                        help='set config keys', default=None,
+                        nargs=argparse.REMAINDER)
+
+  if len(sys.argv) == 1:
+    parser.print_help()
+    sys.exit(1)
+
+  args = parser.parse_args()
+  return args
+
+def combined_roidb(imdb_names):
+  """
+  Combine multiple roidbs
+  """
+
+  def get_roidb(imdb_name):
+    imdb = get_imdb(imdb_name)
+    print('Loaded dataset `{:s}` for training'.format(imdb.name))
+    imdb.set_proposal_method(cfg.TRAIN.PROPOSAL_METHOD)
+    print('Set proposal method: {:s}'.format(cfg.TRAIN.PROPOSAL_METHOD))
+    roidb = get_training_roidb(imdb)
+    return roidb
+
+  roidbs = [get_roidb(s) for s in imdb_names.split('+')]
+  roidb = roidbs[0]
+  if len(roidbs) > 1:
+    for r in roidbs[1:]:
+      roidb.extend(r)
+    tmp = get_imdb(imdb_names.split('+')[1])
+    imdb = datasets.imdb.imdb(imdb_names, tmp.classes)
+  else:
+    imdb = get_imdb(imdb_names)
+  return imdb, roidb
+
+
+if __name__ == '__main__':
+  args = parse_args()
+
+  print('Called with args:')
+  print(args)
+
+  if args.cfg_file is not None:
+    cfg_from_file(args.cfg_file)
+  if args.set_cfgs is not None:
+    cfg_from_list(args.set_cfgs)
+
+  print('Using config:')
+  pprint.pprint(cfg)
+
+  # if has model, get the name from it
+  # if does not, then just use the initialization weights
+  if args.model:
+    filename = os.path.splitext(os.path.basename(args.model))[0]
+  else:
+    filename = os.path.splitext(os.path.basename(args.weight))[0]
+
+  tag = args.tag
+  tag = tag if tag else 'default'
+  filename = tag + '/' + filename
+
+  imdb = get_imdb(args.imdb_name)
+  roidb = imdb.roidb
+  imdb.competition_mode(args.comp_mode)
+
+  # load network
+  if args.net == 'vgg16':
+    net = MELM_vgg16()
+  elif args.net == 'res50':
+    net = resnetv1(num_layers=50)
+  elif args.net == 'res101':
+    net = resnetv1(num_layers=101)
+  elif args.net == 'res152':
+    net = resnetv1(num_layers=152)
+  elif args.net == 'mobile':
+    net = mobilenetv1()
+  else:
+    raise NotImplementedError
+
+  # load model
+  net.create_architecture(imdb.num_classes, tag='default',
+                          anchor_scales=cfg.ANCHOR_SCALES,
+                          anchor_ratios=cfg.ANCHOR_RATIOS)
+
+  net.eval()
+  if not torch.cuda.is_available():
+    net._device = 'cpu'
+  net.to(net._device)
+
+  if args.model:
+    print(('Loading model check point from {:s}').format(args.model))
+    net.load_state_dict(torch.load(args.model, map_location=lambda storage, loc: storage))
+    print('Loaded.')
+  else:
+    print(('Loading initial weights from {:s}').format(args.weight))
+    print('Loaded.')
+
+  test_net(net, imdb, roidb, filename, max_per_image=args.max_per_image)
diff --git a/tools/trainval_net.py b/tools/trainval_net.py
new file mode 100644
index 0000000..240c7f9
--- /dev/null
+++ b/tools/trainval_net.py
@@ -0,0 +1,135 @@
+# --------------------------------------------------------
+# Tensorflow Faster R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Zheqi He, Xinlei Chen, based on code from Ross Girshick
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import _init_paths
+from model.train_val import get_training_roidb, train_net
+from model.config import cfg, cfg_from_file, cfg_from_list, get_output_dir, get_output_tb_dir
+from datasets.factory import get_imdb
+import datasets.imdb
+import argparse
+import pprint
+import numpy as np
+import sys
+import torch
+
+from nets.vgg16 import MELM_vgg16
+
+def parse_args():
+  """
+  Parse input arguments
+  """
+  parser = argparse.ArgumentParser(description='Train a Fast R-CNN network')
+  parser.add_argument('--cfg', dest='cfg_file',
+                      help='optional config file',
+                      default=None, type=str)
+  parser.add_argument('--weight', dest='weight',
+                      help='initialize with pretrained model weights',
+                      type=str)
+  parser.add_argument('--imdb', dest='imdb_name',
+                      help='dataset to train on',
+                      default='voc_2007_trainval', type=str)
+  parser.add_argument('--imdbval', dest='imdbval_name',
+                      help='dataset to validate on',
+                      default='voc_2007_test', type=str)
+  parser.add_argument('--iters', dest='max_iters',
+                      help='number of iterations to train',
+                      default=40000, type=int)
+  parser.add_argument('--tag', dest='tag',
+                      help='tag of the model',
+                      default=None, type=str)
+  parser.add_argument('--net', dest='net',
+                      help='vgg16, res50, res101, res152, mobile',
+                      default='res50', type=str)
+  parser.add_argument('--set', dest='set_cfgs',
+                      help='set config keys', default=None,
+                      nargs=argparse.REMAINDER)
+
+  if len(sys.argv) == 1:
+    parser.print_help()
+    sys.exit(1)
+
+  args = parser.parse_args()
+  return args
+
+
+def combined_roidb(imdb_names):
+  """
+  Combine multiple roidbs
+  """
+
+  def get_roidb(imdb_name):
+    imdb = get_imdb(imdb_name)
+    print('Loaded dataset `{:s}` for training'.format(imdb.name))
+    imdb.set_proposal_method(cfg.TRAIN.PROPOSAL_METHOD)
+    print('Set proposal method: {:s}'.format(cfg.TRAIN.PROPOSAL_METHOD))
+    roidb = get_training_roidb(imdb)
+    return roidb
+
+  roidbs = [get_roidb(s) for s in imdb_names.split('+')]
+  roidb = roidbs[0]
+  if len(roidbs) > 1:
+    for r in roidbs[1:]:
+      roidb.extend(r)
+    tmp = get_imdb(imdb_names.split('+')[1])
+    imdb = datasets.imdb.imdb(imdb_names, tmp.classes)
+  else:
+    imdb = get_imdb(imdb_names)
+  return imdb, roidb
+
+
+if __name__ == '__main__':
+  import os
+  os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+  os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"  # specify which GPU(s) to be used
+
+  args = parse_args()
+  args.set_cfgs = ['ANCHOR_SCALES', '[8,16,32]', 'ANCHOR_RATIOS', '[0.5,1,2]', 'TRAIN.STEPSIZE', '[50000]']
+  print('Called with args:')
+  print(args)
+
+  if args.cfg_file is not None:
+    cfg_from_file(args.cfg_file)
+  if args.set_cfgs is not None:
+    cfg_from_list(args.set_cfgs)
+
+  #print('Using config:')
+  #pprint.pprint(cfg)
+  args.max_iters = 100000
+  np.random.seed(cfg.RNG_SEED)
+  torch.manual_seed(cfg.RNG_SEED)
+  torch.cuda.manual_seed_all(cfg.RNG_SEED)
+
+  # train set
+  imdb, roidb = combined_roidb(args.imdb_name)
+  print('{:d} roidb entries'.format(len(roidb)))
+
+  # output directory where the models are saved
+  output_dir = get_output_dir(imdb, args.tag)
+  print('Output will be saved to `{:s}`'.format(output_dir))
+
+  # tensorboard directory where the summaries are saved during training
+  tb_dir = get_output_tb_dir(imdb, args.tag)
+  print('TensorFlow summaries will be saved to `{:s}`'.format(tb_dir))
+
+  # also add the validation set, but with no flipping images
+  orgflip = cfg.TRAIN.USE_FLIPPED
+  cfg.TRAIN.USE_FLIPPED = False
+  _, valroidb = combined_roidb(args.imdbval_name)
+  print('{:d} validation roidb entries'.format(len(valroidb)))
+  cfg.TRAIN.USE_FLIPPED = orgflip
+
+  # load network
+  if args.net == 'vgg16':
+    net = MELM_vgg16()
+  else:
+    raise NotImplementedError
+    
+  train_net(net, imdb, roidb, valroidb, output_dir, tb_dir,
+            pretrained_model=args.weight,
+            max_iters=args.max_iters)