[Fix] Fix mmdet flops error (open-mmlab#9589)

Author: zwhus

configs/cascade_rcnn/cascade-mask-rcnn_x101-32x8d_fpn_ms-3x_coco.py

@@ -7,7 +7,7 @@
         type='DetDataPreprocessor',
         mean=[103.530, 116.280, 123.675],
         std=[57.375, 57.120, 58.395],
-        to_rgb=False,
+        bgr_to_rgb=False,
         pad_size_divisor=32),
     backbone=dict(
         type='ResNeXt',

configs/cascade_rcnn/cascade-rcnn_r50-caffe_fpn_1x_coco.py

@@ -6,7 +6,7 @@
         type='DetDataPreprocessor',
         mean=[103.530, 116.280, 123.675],
         std=[1.0, 1.0, 1.0],
-        to_rgb=False,
+        bgr_to_rgb=False,
         pad_size_divisor=32),
     backbone=dict(
         norm_cfg=dict(requires_grad=False),

configs/dcnv2/metafile.yml

@@ -19,7 +19,7 @@ Collections:
 Models:
   - Name: faster-rcnn_r50_fpn_mdconv_c3-c5_1x_coco
     In Collection: Deformable Convolutional Networks v2
-    Config: configs/dcn/faster-rcnn_r50-mdconv-c3-c5_fpn_1x_coco.py
+    Config: configs/dcnv2/faster-rcnn_r50-mdconv-c3-c5_fpn_1x_coco.py
     Metadata:
       Training Memory (GB): 4.1
       inference time (ms/im):
@@ -39,7 +39,7 @@ Models:
 
   - Name: faster-rcnn_r50_fpn_mdconv_c3-c5_group4_1x_coco
     In Collection: Deformable Convolutional Networks v2
-    Config: configs/dcn/faster-rcnn_r50-mdconv-group4-c3-c5_fpn_1x_coco.py
+    Config: configs/dcnv2/faster-rcnn_r50-mdconv-group4-c3-c5_fpn_1x_coco.py
     Metadata:
       Training Memory (GB): 4.2
       inference time (ms/im):
@@ -59,7 +59,7 @@ Models:
 
   - Name: faster-rcnn_r50_fpn_mdpool_1x_coco
     In Collection: Deformable Convolutional Networks v2
-    Config: configs/dcn/faster-rcnn_r50_fpn_mdpool_1x_coco.py
+    Config: configs/dcnv2/faster-rcnn_r50_fpn_mdpool_1x_coco.py
     Metadata:
       Training Memory (GB): 5.8
       inference time (ms/im):
@@ -79,7 +79,7 @@ Models:
 
   - Name: mask-rcnn_r50_fpn_mdconv_c3-c5_1x_coco
     In Collection: Deformable Convolutional Networks v2
-    Config: configs/dcn/mask-rcnn_r50-mdconv-c3-c5_fpn_1x_coco.py
+    Config: configs/dcnv2/mask-rcnn_r50-mdconv-c3-c5_fpn_1x_coco.py
     Metadata:
       Training Memory (GB): 4.5
       inference time (ms/im):
@@ -103,7 +103,7 @@ Models:
 
   - Name: mask-rcnn_r50_fpn_fp16_mdconv_c3-c5_1x_coco
     In Collection: Deformable Convolutional Networks v2
-    Config: configs/dcn/mask-rcnn_r50-mdconv-c3-c5_fpn_amp-1x_coco.py
+    Config: configs/dcnv2/mask-rcnn_r50-mdconv-c3-c5_fpn_amp-1x_coco.py
     Metadata:
       Training Memory (GB): 3.1
       Training Techniques:

configs/mask2former/mask2former_r50_8xb2-lsj-50e_coco.py

@@ -1,5 +1,4 @@
 _base_ = ['./mask2former_r50_8xb2-lsj-50e_coco-panoptic.py']
-
 num_things_classes = 80
 num_stuff_classes = 0
 num_classes = num_things_classes + num_stuff_classes
@@ -56,6 +55,17 @@
     dict(type='PackDetInputs')
 ]
 
+test_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=_base_.file_client_args),
+    dict(type='Resize', scale=(1333, 800), keep_ratio=True),
+    # If you don't have a gt annotation, delete the pipeline
+    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
+    dict(
+        type='PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor'))
+]
+
 dataset_type = 'CocoDataset'
 data_root = 'data/coco/'
 
@@ -70,7 +80,8 @@
     dataset=dict(
         type=dataset_type,
         ann_file='annotations/instances_val2017.json',
-        data_prefix=dict(img='val2017/')))
+        data_prefix=dict(img='val2017/'),
+        pipeline=test_pipeline))
 test_dataloader = val_dataloader
 
 val_evaluator = dict(

configs/openimages/faster-rcnn_r50_fpn_32xb2-1x_openimages-challenge.py

@@ -17,6 +17,7 @@
     dataset=dict(
         type=dataset_type,
         ann_file='challenge2019/challenge-2019-validation-detection-bbox.txt',
+        data_prefix=dict(img='OpenImages/'),
         label_file='challenge2019/cls-label-description.csv',
         hierarchy_file='challenge2019/class_label_tree.np',
         meta_file='challenge2019/challenge-2019-validation-metas.pkl',

configs/retinanet/retinanet_r101-caffe_fpn_ms-3x_coco.py

@@ -1,5 +1,8 @@
 _base_ = './retinanet_r50-caffe_fpn_ms-3x_coco.py'
 # learning policy
 model = dict(
-    pretrained='open-mmlab://detectron2/resnet101_caffe',
-    backbone=dict(depth=101))
+    backbone=dict(
+        depth=101,
+        init_cfg=dict(
+            type='Pretrained',
+            checkpoint='open-mmlab://detectron2/resnet101_caffe')))

configs/retinanet/retinanet_r101_fpn_ms-640-800-3x_coco.py

@@ -1,5 +1,9 @@
 _base_ = ['../_base_/models/retinanet_r50_fpn.py', '../common/ms_3x_coco.py']
 # optimizer
-model = dict(pretrained='torchvision://resnet101', backbone=dict(depth=101))
+model = dict(
+    backbone=dict(
+        depth=101,
+        init_cfg=dict(type='Pretrained',
+                      checkpoint='torchvision://resnet101')))
 optim_wrapper = dict(
     optimizer=dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001))

configs/retinanet/retinanet_x101-64x4d_fpn_ms-640-800-3x_coco.py

@@ -1,6 +1,11 @@
 _base_ = ['../_base_/models/retinanet_r50_fpn.py', '../common/ms_3x_coco.py']
 # optimizer
 model = dict(
-    pretrained='open-mmlab://resnext101_64x4d',
-    backbone=dict(type='ResNeXt', depth=101, groups=64, base_width=4))
+    backbone=dict(
+        type='ResNeXt',
+        depth=101,
+        groups=64,
+        base_width=4,
+        init_cfg=dict(
+            type='Pretrained', checkpoint='open-mmlab://resnext101_64x4d')))
 optim_wrapper = dict(optimizer=dict(type='SGD', lr=0.01))

configs/seesaw_loss/metafile.yml

@@ -23,7 +23,7 @@ Collections:
 Models:
   - Name: mask-rcnn_r50_fpn_random_seesaw_loss_mstrain_2x_lvis_v1
     In Collection: Seesaw Loss
-    Config: seesaw_loss/mask-rcnn_r50_fpn_seesaw-loss_random-ms-2x_lvis-v1.py
+    Config: configs/seesaw_loss/mask-rcnn_r50_fpn_seesaw-loss_random-ms-2x_lvis-v1.py
     Metadata:
       Epochs: 24
     Results:
@@ -38,7 +38,7 @@ Models:
     Weights: https://download.openmmlab.com/mmdetection/v2.0/seesaw_loss/mask_rcnn_r50_fpn_random_seesaw_loss_mstrain_2x_lvis_v1-a698dd3d.pth
   - Name: mask-rcnn_r50_fpn_random_seesaw_loss_normed_mask_mstrain_2x_lvis_v1
     In Collection: Seesaw Loss
-    Config: seesaw_loss/mask-rcnn_r50_fpn_seesaw-loss-normed-mask_random-ms-2x_lvis-v1.py
+    Config: configs/seesaw_loss/mask-rcnn_r50_fpn_seesaw-loss-normed-mask_random-ms-2x_lvis-v1.py
     Metadata:
       Epochs: 24
     Results:
@@ -53,7 +53,7 @@ Models:
     Weights: https://download.openmmlab.com/mmdetection/v2.0/seesaw_loss/mask_rcnn_r50_fpn_random_seesaw_loss_normed_mask_mstrain_2x_lvis_v1-a1c11314.pth
   - Name: mask-rcnn_r101_fpn_seesaw-loss_random-ms-2x_lvis-v1
     In Collection: Seesaw Loss
-    Config: seesaw_loss/mask-rcnn_r101_fpn_seesaw-loss_random-ms-2x_lvis-v1.py
+    Config: configs/seesaw_loss/mask-rcnn_r101_fpn_seesaw-loss_random-ms-2x_lvis-v1.py
     Metadata:
       Epochs: 24
     Results:
@@ -68,7 +68,7 @@ Models:
     Weights: https://download.openmmlab.com/mmdetection/v2.0/seesaw_loss/mask_rcnn_r101_fpn_random_seesaw_loss_mstrain_2x_lvis_v1-8e6e6dd5.pth
   - Name: mask-rcnn_r101_fpn_seesaw-loss-normed-mask_random-ms-2x_lvis-v1
     In Collection: Seesaw Loss
-    Config: seesaw_loss/mask-rcnn_r101_fpn_seesaw-loss-normed-mask_random-ms-2x_lvis-v1.py
+    Config: configs/seesaw_loss/mask-rcnn_r101_fpn_seesaw-loss-normed-mask_random-ms-2x_lvis-v1.py
     Metadata:
       Epochs: 24
     Results:

mmdet/models/dense_heads/deformable_detr_head.py

@@ -181,8 +181,7 @@ def forward(self, x: Tuple[Tensor],
                 enc_outputs_class, \
                 enc_outputs_coord.sigmoid()
         else:
-            return outputs_classes, outputs_coords, \
-                None, None
+            return outputs_classes, outputs_coords
 
     def loss_by_feat(
         self,

mmdet/models/detectors/detr.py

@@ -1,5 +1,10 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from typing import List, Tuple
+
+from torch import Tensor
+
 from mmdet.registry import MODELS
+from mmdet.structures import SampleList
 from mmdet.utils import ConfigType, OptConfigType, OptMultiConfig
 from .single_stage import SingleStageDetector
 
@@ -24,3 +29,24 @@ def __init__(self,
             test_cfg=test_cfg,
             data_preprocessor=data_preprocessor,
             init_cfg=init_cfg)
+
+    def _forward(self, batch_inputs: Tensor,
+                 batch_data_samples: SampleList) -> Tuple[List[Tensor]]:
+        """Network forward process. Usually includes backbone, neck and head
+        forward without any post-processing.
+
+         Args:
+            batch_inputs (Tensor): Inputs with shape (N, C, H, W).
+            batch_data_samples (list[:obj:`DetDataSample`]): The batch
+                data samples. It usually includes information such
+                as `gt_instance` or `gt_panoptic_seg` or `gt_sem_seg`.
+
+        Returns:
+            tuple[list]: A tuple of features from ``bbox_head`` forward.
+        """
+        x = self.extract_feat(batch_inputs)
+        batch_img_metas = [
+            data_samples.metainfo for data_samples in batch_data_samples
+        ]
+        results = self.bbox_head.forward(x, batch_img_metas)
+        return results

mmdet/models/detectors/single_stage.py

@@ -122,6 +122,9 @@ def _forward(
 
          Args:
             batch_inputs (Tensor): Inputs with shape (N, C, H, W).
+            batch_data_samples (list[:obj:`DetDataSample`]): Each item contains
+                the meta information of each image and corresponding
+                annotations.
 
         Returns:
             tuple[list]: A tuple of features from ``bbox_head`` forward.

mmdet/models/detectors/two_stage.py

@@ -119,6 +119,9 @@ def _forward(self, batch_inputs: Tensor,
 
         Args:
             batch_inputs (Tensor): Inputs with shape (N, C, H, W).
+            batch_data_samples (list[:obj:`DetDataSample`]): Each item contains
+                the meta information of each image and corresponding
+                annotations.
 
         Returns:
             tuple: A tuple of features from ``rpn_head`` and ``roi_head``
@@ -135,8 +138,8 @@ def _forward(self, batch_inputs: Tensor,
             rpn_results_list = [
                 data_sample.proposals for data_sample in batch_data_samples
             ]
-
-        roi_outs = self.roi_head.forward(x, rpn_results_list)
+        roi_outs = self.roi_head.forward(x, rpn_results_list,
+                                         batch_data_samples)
         results = results + (roi_outs, )
         return results
 

mmdet/models/roi_heads/cascade_roi_head.py

@@ -518,3 +518,51 @@ def _refine_roi(self, x: Tuple[Tensor], rois: Tensor,
             for i in range(len(batch_img_metas))
         ]
         return rois, cls_scores, bbox_preds
+
+    def forward(self, x: Tuple[Tensor], rpn_results_list: InstanceList,
+                batch_data_samples: SampleList) -> tuple:
+        """Network forward process. Usually includes backbone, neck and head
+        forward without any post-processing.
+
+        Args:
+            x (List[Tensor]): Multi-level features that may have different
+                resolutions.
+            rpn_results_list (list[:obj:`InstanceData`]): List of region
+                proposals.
+            batch_data_samples (list[:obj:`DetDataSample`]): Each item contains
+                the meta information of each image and corresponding
+                annotations.
+
+        Returns
+            tuple: A tuple of features from ``bbox_head`` and ``mask_head``
+            forward.
+        """
+        results = ()
+        batch_img_metas = [
+            data_samples.metainfo for data_samples in batch_data_samples
+        ]
+        proposals = [rpn_results.bboxes for rpn_results in rpn_results_list]
+        num_proposals_per_img = tuple(len(p) for p in proposals)
+        rois = bbox2roi(proposals)
+        # bbox head
+        if self.with_bbox:
+            rois, cls_scores, bbox_preds = self._refine_roi(
+                x, rois, batch_img_metas, num_proposals_per_img)
+            results = results + (cls_scores, bbox_preds)
+        # mask head
+        if self.with_mask:
+            aug_masks = []
+            rois = torch.cat(rois)
+            for stage in range(self.num_stages):
+                mask_results = self._mask_forward(stage, x, rois)
+                mask_preds = mask_results['mask_preds']
+                mask_preds = mask_preds.split(num_proposals_per_img, 0)
+                aug_masks.append([m.sigmoid().detach() for m in mask_preds])
+
+            merged_masks = []
+            for i in range(len(batch_img_metas)):
+                aug_mask = [mask[i] for mask in aug_masks]
+                merged_mask = merge_aug_masks(aug_mask, batch_img_metas[i])
+                merged_masks.append(merged_mask)
+            results = results + (merged_masks, )
+        return results

mmdet/models/roi_heads/grid_roi_head.py

@@ -75,8 +75,10 @@ def _random_jitter(self,
         return sampling_results
 
     # TODO: Forward is incorrect and need to refactor.
-    def forward(self, x: Tuple[Tensor],
-                rpn_results_list: InstanceList) -> tuple:
+    def forward(self,
+                x: Tuple[Tensor],
+                rpn_results_list: InstanceList,
+                batch_data_samples: SampleList = None) -> tuple:
         """Network forward process. Usually includes backbone, neck and head
         forward without any post-processing.
 
@@ -85,6 +87,9 @@ def forward(self, x: Tuple[Tensor],
                 resolutions.
             rpn_results_list (list[:obj:`InstanceData`]): List of region
                 proposals.
+            batch_data_samples (list[:obj:`DetDataSample`]): Each item contains
+            the meta information of each image and corresponding
+            annotations.
 
         Returns
             tuple: A tuple of features from ``bbox_head`` and ``mask_head``
@@ -96,8 +101,9 @@ def forward(self, x: Tuple[Tensor],
         # bbox head
         if self.with_bbox:
             bbox_results = self._bbox_forward(x, rois)
-            results = results + (bbox_results['cls_score'],
-                                 bbox_results['bbox_pred'])
+            results = results + (bbox_results['cls_score'], )
+            if self.bbox_head.with_reg:
+                results = results + (bbox_results['bbox_pred'], )
 
             # grid head
             grid_rois = rois[:100]