Merge branch 'dev' of https://github.com/open-mmlab/mmdetection into …

…multi_machine_dist

.github/workflows/build.yml

@@ -265,7 +265,7 @@ jobs:
       - name: Build and install
         run: pip install -e .
       - name: Run unittests
-        run: coverage run --branch --source mmdet -m pytest tests -sv
+        run: coverage run --branch --source mmdet -m pytest tests
       - name: Generate coverage report
         run: |
           coverage xml

configs/yolox/yolox_s_8x8_300e_coco.py

@@ -1,6 +1,6 @@
 _base_ = ['../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py']
 
-img_scale = (640, 640)
+img_scale = (640, 640)  # height, width
 
 # model settings
 model = dict(

configs/yolox/yolox_tiny_8x8_300e_coco.py

@@ -7,7 +7,7 @@
     neck=dict(in_channels=[96, 192, 384], out_channels=96),
     bbox_head=dict(in_channels=96, feat_channels=96))
 
-img_scale = (640, 640)
+img_scale = (640, 640)  # height, width
 
 train_pipeline = [
     dict(type='Mosaic', img_scale=img_scale, pad_val=114.0),

docs/zh_cn/tutorials/finetune.md

@@ -1 +1,84 @@
 # 教程 7: 模型微调
+
+在 COCO 数据集上预训练的检测器可以作为其他数据集（例如 CityScapes 和 KITTI 数据集）优质的预训练模型。
+本教程将指导用户如何把 [ModelZoo](../model_zoo.md) 中提供的模型用于其他数据集中并使得当前所训练的模型获得更好性能。
+
+以下是在新数据集中微调模型需要的两个步骤。
+
+- 按 [教程2：自定义数据集的方法](customize_dataset.md) 中的方法对新数据集添加支持中的方法对新数据集添加支持
+- 按照本教程中所讨论方法，修改配置信息
+
+接下来将会以 Cityscapes Dataset 上的微调过程作为例子，具体讲述用户需要在配置中修改的五个部分。
+
+## 继承基础配置
+
+为了减轻编写整个配置的负担并减少漏洞的数量， MMDetection V2.0 支持从多个现有配置中继承配置信息。微调 MaskRCNN 模型的时候，新的配置信息需要使用从 `_base_/models/mask_rcnn_r50_fpn.py`中继承的配置信息来构建模型的基本结构。当使用 Cityscapes 数据集时，新的配置信息可以简便地从`_base_/datasets/cityscapes_instance.py`中继承。对于训练过程的运行设置部分，新配置需要从 `_base_/default_runtime.py`中继承。这些配置文件`configs`的目录下，用户可以选择全部内容的重新编写而不是使用继承方法。
+
+```python
+_base_ = [
+    '../_base_/models/mask_rcnn_r50_fpn.py',
+    '../_base_/datasets/cityscapes_instance.py', '../_base_/default_runtime.py'
+]
+```
+
+
+##  Head 的修改
+接下来新的配置还需要根据新数据集的类别数量对 Head 进行修改。只需要对 roi_head 中的 `num_classes`进行修改。修改后除了最后的预测模型的 Head 之外，预训练模型的权重的大部分都会被重新使用。
+
+```python
+model = dict(
+    pretrained=None,
+    roi_head=dict(
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=256,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=8,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
+        mask_head=dict(
+            type='FCNMaskHead',
+            num_convs=4,
+            in_channels=256,
+            conv_out_channels=256,
+            num_classes=8,
+            loss_mask=dict(
+                type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
+```
+
+## 数据集的修改
+用户可能还需要准备数据集并编写有关数据集的配置。目前 MMDetection V2.0 的配置文件已经支持 VOC、WIDER FACE、COCO 和 Cityscapes Dataset 的数据集信息。
+
+## 训练策略的修改
+微调超参数与默认的训练策略不同。它通常需要更小的学习率和更少的训练回合。
+
+```python
+# 优化器
+# batch size 为 8 时的 lr 配置
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
+optimizer_config = dict(grad_clip=None)
+# 学习策略
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[7])
+# lr_config 中的 max_epochs 和 step 需要针对自定义数据集进行专门调整
+runner = dict(max_epochs=8)
+log_config = dict(interval=100)
+```
+
+## 使用预训练模型
+
+如果要使用预训练模型时，可以在 `load_from` 中查阅新的配置信息，用户需要在训练开始之前下载好需要的模型权重，从而避免在训练过程中浪费了宝贵时间。
+```python
+load_from = 'https://download.openmmlab.com/mmdetection/v2.0/mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth'  # noqa
+```

mmdet/datasets/api_wrappers/panoptic_evaluation.py

@@ -170,7 +170,8 @@ def pq_compute_multi_core(matched_annotations_list,
                           gt_folder,
                           pred_folder,
                           categories,
-                          file_client=None):
+                          file_client=None,
+                          nproc=32):
     """Evaluate the metrics of Panoptic Segmentation with multithreading.
 
     Same as the function with the same name in `panopticapi`.
@@ -184,6 +185,9 @@ def pq_compute_multi_core(matched_annotations_list,
         categories (str): The categories of the dataset.
         file_client (object): The file client of the dataset. If None,
             the backend will be set to `disk`.
+        nproc (int): Number of processes for panoptic quality computing.
+            Defaults to 32. When `nproc` exceeds the number of cpu cores,
+            the number of cpu cores is used.
     """
     if PQStat is None:
         raise RuntimeError(
@@ -195,7 +199,8 @@ def pq_compute_multi_core(matched_annotations_list,
         file_client_args = dict(backend='disk')
         file_client = mmcv.FileClient(**file_client_args)
 
-    cpu_num = multiprocessing.cpu_count()
+    cpu_num = min(nproc, multiprocessing.cpu_count())
+
     annotations_split = np.array_split(matched_annotations_list, cpu_num)
     print('Number of cores: {}, images per core: {}'.format(
         cpu_num, len(annotations_split[0])))

mmdet/datasets/coco_panoptic.py

@@ -426,7 +426,8 @@ def evaluate_pan_json(self,
                           result_files,
                           outfile_prefix,
                           logger=None,
-                          classwise=False):
+                          classwise=False,
+                          nproc=32):
         """Evaluate PQ according to the panoptic results json file."""
         imgs = self.coco.imgs
         gt_json = self.coco.img_ann_map  # image to annotations
@@ -451,9 +452,13 @@ def evaluate_pan_json(self,
         gt_folder = self.seg_prefix
         pred_folder = os.path.join(os.path.dirname(outfile_prefix), 'panoptic')
 
-        pq_stat = pq_compute_multi_core(matched_annotations_list, gt_folder,
-                                        pred_folder, self.categories,
-                                        self.file_client)
+        pq_stat = pq_compute_multi_core(
+            matched_annotations_list,
+            gt_folder,
+            pred_folder,
+            self.categories,
+            self.file_client,
+            nproc=nproc)
 
         metrics = [('All', None), ('Things', True), ('Stuff', False)]
         pq_results = {}
@@ -480,6 +485,7 @@ def evaluate(self,
                  logger=None,
                  jsonfile_prefix=None,
                  classwise=False,
+                 nproc=32,
                  **kwargs):
         """Evaluation in COCO Panoptic protocol.
 
@@ -494,6 +500,9 @@ def evaluate(self,
                 If not specified, a temp file will be created. Default: None.
             classwise (bool): Whether to print classwise evaluation results.
                 Default: False.
+            nproc (int): Number of processes for panoptic quality computing.
+                Defaults to 32. When `nproc` exceeds the number of cpu cores,
+                the number of cpu cores is used.
 
         Returns:
             dict[str, float]: COCO Panoptic style evaluation metric.
@@ -512,9 +521,8 @@ def evaluate(self,
         outfile_prefix = os.path.join(tmp_dir.name, 'results') \
             if tmp_dir is not None else jsonfile_prefix
         if 'PQ' in metrics:
-            eval_pan_results = self.evaluate_pan_json(result_files,
-                                                      outfile_prefix, logger,
-                                                      classwise)
+            eval_pan_results = self.evaluate_pan_json(
+                result_files, outfile_prefix, logger, classwise, nproc=nproc)
             eval_results.update(eval_pan_results)
 
         if tmp_dir is not None:

mmdet/datasets/custom.py

@@ -74,8 +74,8 @@ def __init__(self,
         self.proposal_file = proposal_file
         self.test_mode = test_mode
         self.filter_empty_gt = filter_empty_gt
-        self.CLASSES = self.get_classes(classes)
         self.file_client = mmcv.FileClient(**file_client_args)
+        self.CLASSES = self.get_classes(classes)
 
         # join paths if data_root is specified
         if self.data_root is not None:

mmdet/datasets/pipelines/loading.py

@@ -37,9 +37,11 @@ class LoadImageFromFile:
     def __init__(self,
                  to_float32=False,
                  color_type='color',
+                 channel_order='bgr',
                  file_client_args=dict(backend='disk')):
         self.to_float32 = to_float32
         self.color_type = color_type
+        self.channel_order = channel_order
         self.file_client_args = file_client_args.copy()
         self.file_client = None
 
@@ -63,7 +65,8 @@ def __call__(self, results):
             filename = results['img_info']['filename']
 
         img_bytes = self.file_client.get(filename)
-        img = mmcv.imfrombytes(img_bytes, flag=self.color_type)
+        img = mmcv.imfrombytes(
+            img_bytes, flag=self.color_type, channel_order=self.channel_order)
         if self.to_float32:
             img = img.astype(np.float32)
 
@@ -79,6 +82,7 @@ def __repr__(self):
         repr_str = (f'{self.__class__.__name__}('
                     f'to_float32={self.to_float32}, '
                     f"color_type='{self.color_type}', "
+                    f"channel_order='{self.channel_order}', "
                     f'file_client_args={self.file_client_args})')
         return repr_str
 
@@ -438,9 +442,14 @@ def __init__(self,
                 'pip install git+https://github.com/cocodataset/'
                 'panopticapi.git.')
 
-        super(LoadPanopticAnnotations,
-              self).__init__(with_bbox, with_label, with_mask, with_seg, True,
-                             file_client_args)
+        super(LoadPanopticAnnotations, self).__init__(
+            with_bbox=with_bbox,
+            with_label=with_label,
+            with_mask=with_mask,
+            with_seg=with_seg,
+            poly2mask=True,
+            denorm_bbox=False,
+            file_client_args=file_client_args)
 
     def _load_masks_and_semantic_segs(self, results):
         """Private function to load mask and semantic segmentation annotations.

mmdet/datasets/pipelines/transforms.py

@@ -11,6 +11,7 @@
 
 from mmdet.core import PolygonMasks, find_inside_bboxes
 from mmdet.core.evaluation.bbox_overlaps import bbox_overlaps
+from mmdet.utils import log_img_scale
 from ..builder import PIPELINES
 
 try:
@@ -1979,9 +1980,10 @@ class Mosaic:
 
     Args:
         img_scale (Sequence[int]): Image size after mosaic pipeline of single
-           image. Default to (640, 640).
+            image. The shape order should be (height, width).
+            Default to (640, 640).
         center_ratio_range (Sequence[float]): Center ratio range of mosaic
-           output. Default to (0.5, 1.5).
+            output. Default to (0.5, 1.5).
         min_bbox_size (int | float): The minimum pixel for filtering
             invalid bboxes after the mosaic pipeline. Default to 0.
         bbox_clip_border (bool, optional): Whether to clip the objects outside
@@ -2002,6 +2004,7 @@ def __init__(self,
                  skip_filter=True,
                  pad_val=114):
         assert isinstance(img_scale, tuple)
+        log_img_scale(img_scale, skip_square=True)
         self.img_scale = img_scale
         self.center_ratio_range = center_ratio_range
         self.min_bbox_size = min_bbox_size
@@ -2232,7 +2235,7 @@ class MixUp:
                 |             pad              |
                 +------------------------------+
 
-     The mixup transform steps are as follows::
+     The mixup transform steps are as follows:
 
         1. Another random image is picked by dataset and embedded in
            the top left patch(after padding and resizing)
@@ -2241,15 +2244,15 @@ class MixUp:
 
     Args:
         img_scale (Sequence[int]): Image output size after mixup pipeline.
-           Default: (640, 640).
+            The shape order should be (height, width). Default: (640, 640).
         ratio_range (Sequence[float]): Scale ratio of mixup image.
-           Default: (0.5, 1.5).
+            Default: (0.5, 1.5).
         flip_ratio (float): Horizontal flip ratio of mixup image.
-           Default: 0.5.
+            Default: 0.5.
         pad_val (int): Pad value. Default: 114.
         max_iters (int): The maximum number of iterations. If the number of
-           iterations is greater than `max_iters`, but gt_bbox is still
-           empty, then the iteration is terminated. Default: 15.
+            iterations is greater than `max_iters`, but gt_bbox is still
+            empty, then the iteration is terminated. Default: 15.
         min_bbox_size (float): Width and height threshold to filter bboxes.
             If the height or width of a box is smaller than this value, it
             will be removed. Default: 5.
@@ -2281,6 +2284,7 @@ def __init__(self,
                  bbox_clip_border=True,
                  skip_filter=True):
         assert isinstance(img_scale, tuple)
+        log_img_scale(img_scale, skip_square=True)
         self.dynamic_scale = img_scale
         self.ratio_range = ratio_range
         self.flip_ratio = flip_ratio

mmdet/models/dense_heads/centripetal_head.py

@@ -2,6 +2,7 @@
 import torch.nn as nn
 from mmcv.cnn import ConvModule, normal_init
 from mmcv.ops import DeformConv2d
+from mmcv.runner import force_fp32
 
 from mmdet.core import multi_apply
 from ..builder import HEADS, build_loss
@@ -203,6 +204,7 @@ def forward_single(self, x, lvl_ind):
         ]
         return result_list
 
+    @force_fp32()
     def loss(self,
              tl_heats,
              br_heats,
@@ -361,6 +363,7 @@ def loss_single(self, tl_hmp, br_hmp, tl_off, br_off, tl_guiding_shift,
 
         return det_loss, off_loss, guiding_loss, centripetal_loss
 
+    @force_fp32()
     def get_bboxes(self,
                    tl_heats,
                    br_heats,

mmdet/models/dense_heads/corner_head.py

@@ -6,7 +6,7 @@
 import torch.nn as nn
 from mmcv.cnn import ConvModule, bias_init_with_prob
 from mmcv.ops import CornerPool, batched_nms
-from mmcv.runner import BaseModule
+from mmcv.runner import BaseModule, force_fp32
 
 from mmdet.core import multi_apply
 from ..builder import HEADS, build_loss
@@ -152,6 +152,7 @@ def __init__(self,
         self.train_cfg = train_cfg
         self.test_cfg = test_cfg
 
+        self.fp16_enabled = False
         self._init_layers()
 
     def _make_layers(self, out_channels, in_channels=256, feat_channels=256):
@@ -509,6 +510,7 @@ def get_targets(self,
 
         return target_result
 
+    @force_fp32()
     def loss(self,
              tl_heats,
              br_heats,
@@ -649,6 +651,7 @@ def loss_single(self, tl_hmp, br_hmp, tl_emb, br_emb, tl_off, br_off,
 
         return det_loss, pull_loss, push_loss, off_loss
 
+    @force_fp32()
     def get_bboxes(self,
                    tl_heats,
                    br_heats,

mmdet/models/dense_heads/ssd_head.py

@@ -316,7 +316,7 @@ def loss(self,
             gt_bboxes_ignore_list=gt_bboxes_ignore,
             gt_labels_list=gt_labels,
             label_channels=1,
-            unmap_outputs=False)
+            unmap_outputs=True)
         if cls_reg_targets is None:
             return None
         (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,

mmdet/models/dense_heads/yolox_head.py

@@ -212,6 +212,7 @@ def forward(self, feats):
                            self.multi_level_conv_reg,
                            self.multi_level_conv_obj)
 
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'objectnesses'))
     def get_bboxes(self,
                    cls_scores,
                    bbox_preds,