[Fix] Fix bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py

open-mmlab · Aug 9, 2022 · 4f42839 · 4f42839
1 parent 4eaa8e6
commit 4f42839
Show file tree

Hide file tree

Showing 3 changed files with 65 additions and 5 deletions.
diff --git a/configs/bisenetv2/README.md b/configs/bisenetv2/README.md
@@ -42,7 +42,7 @@ The low-level details and high-level semantics are both essential to the semanti
 | Method           | Backbone  | Crop Size | Lr schd | Mem (GB) | Inf time (fps) |  mIoU | mIoU(ms+flip) | config                                                                                                                                   | download                                                                                                                                                                                                                                                                                                                                                                                               |
 | ---------------- | --------- | --------- | ------: | -------- | -------------- | ----: | ------------: | ---------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
 | BiSeNetV2        | BiSeNetV2 | 1024x1024 |  160000 | 7.64     | 31.77          | 73.21 |         75.74 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes.py)      | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes_20210902_015551-bcf10f09.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes_20210902_015551.log.json)                     |
-| BiSeNetV2 (OHEM) | BiSeNetV2 | 1024x1024 |  160000 | 7.64     | -              | 73.57 |         75.80 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947-5f8103b4.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947.log.json) |
+| BiSeNetV2 (OHEM) | BiSeNetV2 | 1024x1024 |  160000 | 7.64     | -              | 75.30 |         77.06 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324-8bf0aaba.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324.log.json) |
 | BiSeNetV2 (4x8)  | BiSeNetV2 | 1024x1024 |  160000 | 15.05    | -              | 75.76 |         77.79 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes.py)      | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes_20210903_000032-e1a2eed6.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes_20210903_000032.log.json)                     |
 | BiSeNetV2 (FP16) | BiSeNetV2 | 1024x1024 |  160000 | 5.77     | 36.65          | 73.07 |         75.13 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes_20210902_045942-b979777b.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes_20210902_045942.log.json) |
 

diff --git a/configs/bisenetv2/bisenetv2.yml b/configs/bisenetv2/bisenetv2.yml
@@ -45,10 +45,10 @@ Models:
   - Task: Semantic Segmentation
     Dataset: Cityscapes
     Metrics:
-      mIoU: 73.57
-      mIoU(ms+flip): 75.8
+      mIoU: 75.3
+      mIoU(ms+flip): 77.06
   Config: configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py
-  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947-5f8103b4.pth
+  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324-8bf0aaba.pth
 - Name: bisenetv2_fcn_4x8_1024x1024_160k_cityscapes
   In Collection: BiSeNetV2
   Metadata:

diff --git a/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py b/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py
@@ -3,7 +3,67 @@
     '../_base_/datasets/cityscapes_1024x1024.py',
     '../_base_/default_runtime.py', '../_base_/schedules/schedule_160k.py'
 ]
-sampler = dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)
+# sampler = dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    decode_head=dict(
+        sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)),
+    auxiliary_head=[
+        dict(
+            type='FCNHead',
+            in_channels=16,
+            channels=16,
+            num_convs=2,
+            num_classes=19,
+            in_index=1,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+        dict(
+            type='FCNHead',
+            in_channels=32,
+            channels=64,
+            num_convs=2,
+            num_classes=19,
+            in_index=2,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+        dict(
+            type='FCNHead',
+            in_channels=64,
+            channels=256,
+            num_convs=2,
+            num_classes=19,
+            in_index=3,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+        dict(
+            type='FCNHead',
+            in_channels=128,
+            channels=1024,
+            num_convs=2,
+            num_classes=19,
+            in_index=4,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    ],
+)
+
 lr_config = dict(warmup='linear', warmup_iters=1000)
 optimizer = dict(lr=0.05)
 data = dict(