[中国海洋大学][高校贡献][Pytorch迁移1.8][SOLOv2]-初次提交-ModelZoo-PyTorch-AtomGit

MPyTorch/contrib/cv/detection/SOLOv2/README.md+30-7

	@@ -1,19 +1,21 @@
1	# SOLOv2	1	# SOLOv2
2		2
3	-This implements training of SOLOv2 on the Coco dataset, mainly modified from [pytorch/examples](https://github.com/WXinlong/SOLO).	3	+This implements training of SOLOv2 on the Coco dataset, mainly modified
		4	+from [pytorch/examples](https://github.com/WXinlong/SOLO).
4		5
5	## SOLOv2 Detail	6	## SOLOv2 Detail
6		7
7	-As of the current date, Ascend-Pytorch is still inefficient for contiguous operations.	8	+As of the current date, Ascend-Pytorch is still inefficient for contiguous operations. Therefore, SOLOv2 model need to
8	-Therefore, SOLOv2 model need to be modified in the following aspects:	9	+be modified in the following aspects:
9		10
10	1. Converting tensors with the dynamic shapes into tensors with fixed shapes. (This is the hardest one)	11	1. Converting tensors with the dynamic shapes into tensors with fixed shapes. (This is the hardest one)
11	-2. Several operations, like the sum of `INT64`, are not supported on the NPU, so we modified tensors' `dtype` when needed	12	+2. Several operations, like the sum of `INT64`, are not supported on the NPU, so we modified tensors' `dtype` when
12	-3. Framework bottlenecks lead to poor performance, so we improve the original code to improve the performance of the model	13	+ needed
		14	+3. Framework bottlenecks lead to poor performance, so we improve the original code to improve the performance of the
		15	+ model
13	4. We used Apex for mmdtection due to the hardware defects of the NPU	16	4. We used Apex for mmdtection due to the hardware defects of the NPU
14	5. ...	17	5. ...
15		18
16	-
17	## Requirements	19	## Requirements
18		20
19	- NPU配套的run包安装	21	- NPU配套的run包安装
	@@ -21,11 +23,14 @@ Therefore, SOLOv2 model need to be modified in the following aspects:
21	- PyTorch(NPU版本)	23	- PyTorch(NPU版本)
22	- apex(NPU版本)	24	- apex(NPU版本)
23	- MMCV v0.2.16	25	- MMCV v0.2.16
		26	+
24	### Document and data preparation	27	### Document and data preparation
		28	+
25	1. 下载压缩modelzoo\contrib\PyTorch\cv\instance_segmentation\SOLOv2 文件夹	29	1. 下载压缩modelzoo\contrib\PyTorch\cv\instance_segmentation\SOLOv2 文件夹
26	2. 于npu服务器解压SOLOv2压缩包	30	2. 于npu服务器解压SOLOv2压缩包
27	3. 下载coco2017数据集	31	3. 下载coco2017数据集
28	4. 将coco数据集放于SOLOv2/data目录下，目录结构如下：	32	4. 将coco数据集放于SOLOv2/data目录下，目录结构如下：
		33	+
29	```	34	```
30	GFocalV2	35	GFocalV2
31	├── configs	36	├── configs
	@@ -35,13 +40,17 @@ GFocalV2
35	│ ├── train2017 19G	40	│ ├── train2017 19G
36	│ ├── val2017 788M	41	│ ├── val2017 788M
37	```	42	```
		43	+
38	### Configure the environment	44	### Configure the environment
		45	+
39	```	46	```
40	进入SOLOv2目录，source环境变量	47	进入SOLOv2目录，source环境变量
41	cd SOLOv2	48	cd SOLOv2
42	source test/env_npu.sh	49	source test/env_npu.sh
43	```	50	```
		51	+
44	1. 配置安装mmcv	52	1. 配置安装mmcv
		53	+
45	```	54	```
46	cd mmcv	55	cd mmcv
47	python3.7 setup.py build_ext	56	python3.7 setup.py build_ext
	@@ -49,42 +58,56 @@ python3.7 setup.py develop
49	cd ..	58	cd ..
50	pip list \| grep mmcv # 查看版本和路径	59	pip list \| grep mmcv # 查看版本和路径
51	```	60	```
		61	+
52	2. 配置安装mmdet	62	2. 配置安装mmdet
		63	+
53	```	64	```
54	pip install -r requirements/build.txt	65	pip install -r requirements/build.txt
55	pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"	66	pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
56	pip install -v -e .	67	pip install -v -e .
57	```	68	```
		69	+
58	## Train MODEL	70	## Train MODEL
		71	+
59	进入SOLOv2目录下	72	进入SOLOv2目录下
		73	+
60	### 1p	74	### 1p
		75	+
61	导入环境变量，修改train_full_1p.sh权限并运行	76	导入环境变量，修改train_full_1p.sh权限并运行
		77	+
62	```	78	```
63	chmod +x ./test/train_full_1p.sh	79	chmod +x ./test/train_full_1p.sh
64	bash ./test/train_full_1p.sh --data_path=./data/coco	80	bash ./test/train_full_1p.sh --data_path=./data/coco
65	```	81	```
66		82
67	### 8p	83	### 8p
		84	+
68	导入环境变量，修改train_full_8p.sh权限并运行	85	导入环境变量，修改train_full_8p.sh权限并运行
		86	+
69	```	87	```
70	chmod +x ./test/train_full_8p.sh	88	chmod +x ./test/train_full_8p.sh
71	bash ./test/train_full_8p.sh --data_path=./data/coco	89	bash ./test/train_full_8p.sh --data_path=./data/coco
72	```	90	```
73		91
74	### Eval	92	### Eval
		93	+
75	修改train_eval_1p.sh权限并运行	94	修改train_eval_1p.sh权限并运行
		95	+
76	```	96	```
77	chmod +x ./test/train_eval_1p.sh	97	chmod +x ./test/train_eval_1p.sh
78	bash ./test/train_eval_1p.sh --data_path=./data/coco	98	bash ./test/train_eval_1p.sh --data_path=./data/coco
79	```	99	```
		100	+
80	### finetuning	101	### finetuning
		102	+
81	修改train_finetune_1p.sh权限并运行	103	修改train_finetune_1p.sh权限并运行
		104	+
82	```	105	```
83	chmod +x ./test/train_eval_1p.sh	106	chmod +x ./test/train_eval_1p.sh
84	bash ./test/train_finetune_1p.sh --data_path=./data/coco	107	bash ./test/train_finetune_1p.sh --data_path=./data/coco
85	```	108	```
86		109
87	-## SOLOv2 training result	110	+## SOLOv2 training result
88		111
89	\| Acc@1 \| FPS \| Npu/Gpu_nums \| Epochs \| AMP_Type \| Loss_Scale \|	112	\| Acc@1 \| FPS \| Npu/Gpu_nums \| Epochs \| AMP_Type \| Loss_Scale \|
90	\| :------: \| :------: \| :------: \| :------: \| :------: \| :------: \|	113	\| :------: \| :------: \| :------: \| :------: \| :------: \| :------: \|

MPyTorch/contrib/cv/detection/SOLOv2/configs/cascade_mask_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -24,7 +24,7 @@ model = dict(
24	num_stages=3,	24	num_stages=3,
25	strides=(1, 2, 2),	25	strides=(1, 2, 2),
26	dilations=(1, 1, 1),	26	dilations=(1, 1, 1),
27	- out_indices=(2, ),	27	+ out_indices=(2,),
28	frozen_stages=1,	28	frozen_stages=1,
29	norm_cfg=norm_cfg,	29	norm_cfg=norm_cfg,
30	norm_eval=True,	30	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/cascade_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -24,7 +24,7 @@ model = dict(
24	num_stages=3,	24	num_stages=3,
25	strides=(1, 2, 2),	25	strides=(1, 2, 2),
26	dilations=(1, 1, 1),	26	dilations=(1, 1, 1),
27	- out_indices=(2, ),	27	+ out_indices=(2,),
28	frozen_stages=1,	28	frozen_stages=1,
29	norm_cfg=norm_cfg,	29	norm_cfg=norm_cfg,
30	norm_eval=True,	30	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py+3-3

	@@ -151,19 +151,19 @@ data = dict(
151	dataset=dict(	151	dataset=dict(
152	type=dataset_type,	152	type=dataset_type,
153	ann_file=data_root +	153	ann_file=data_root +
154	- 'annotations/instancesonly_filtered_gtFine_train.json',	154	+ 'annotations/instancesonly_filtered_gtFine_train.json',
155	img_prefix=data_root + 'train/',	155	img_prefix=data_root + 'train/',
156	pipeline=train_pipeline)),	156	pipeline=train_pipeline)),
157	val=dict(	157	val=dict(
158	type=dataset_type,	158	type=dataset_type,
159	ann_file=data_root +	159	ann_file=data_root +
160	- 'annotations/instancesonly_filtered_gtFine_val.json',	160	+ 'annotations/instancesonly_filtered_gtFine_val.json',
161	img_prefix=data_root + 'val/',	161	img_prefix=data_root + 'val/',
162	pipeline=test_pipeline),	162	pipeline=test_pipeline),
163	test=dict(	163	test=dict(
164	type=dataset_type,	164	type=dataset_type,
165	ann_file=data_root +	165	ann_file=data_root +
166	- 'annotations/instancesonly_filtered_gtFine_val.json',	166	+ 'annotations/instancesonly_filtered_gtFine_val.json',
167	img_prefix=data_root + 'val/',	167	img_prefix=data_root + 'val/',
168	pipeline=test_pipeline))	168	pipeline=test_pipeline))
169	# optimizer	169	# optimizer

MPyTorch/contrib/cv/detection/SOLOv2/configs/cityscapes/mask_rcnn_r50_fpn_1x_cityscapes.py+3-3

	@@ -165,19 +165,19 @@ data = dict(
165	dataset=dict(	165	dataset=dict(
166	type=dataset_type,	166	type=dataset_type,
167	ann_file=data_root +	167	ann_file=data_root +
168	- 'annotations/instancesonly_filtered_gtFine_train.json',	168	+ 'annotations/instancesonly_filtered_gtFine_train.json',
169	img_prefix=data_root + 'train/',	169	img_prefix=data_root + 'train/',
170	pipeline=train_pipeline)),	170	pipeline=train_pipeline)),
171	val=dict(	171	val=dict(
172	type=dataset_type,	172	type=dataset_type,
173	ann_file=data_root +	173	ann_file=data_root +
174	- 'annotations/instancesonly_filtered_gtFine_val.json',	174	+ 'annotations/instancesonly_filtered_gtFine_val.json',
175	img_prefix=data_root + 'val/',	175	img_prefix=data_root + 'val/',
176	pipeline=test_pipeline),	176	pipeline=test_pipeline),
177	test=dict(	177	test=dict(
178	type=dataset_type,	178	type=dataset_type,
179	ann_file=data_root +	179	ann_file=data_root +
180	- 'annotations/instancesonly_filtered_gtFine_val.json',	180	+ 'annotations/instancesonly_filtered_gtFine_val.json',
181	img_prefix=data_root + 'val/',	181	img_prefix=data_root + 'val/',
182	pipeline=test_pipeline))	182	pipeline=test_pipeline))
183	# optimizer	183	# optimizer

MPyTorch/contrib/cv/detection/SOLOv2/configs/fast_mask_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -23,7 +23,7 @@ model = dict(
23	num_stages=3,	23	num_stages=3,
24	strides=(1, 2, 2),	24	strides=(1, 2, 2),
25	dilations=(1, 1, 1),	25	dilations=(1, 1, 1),
26	- out_indices=(2, ),	26	+ out_indices=(2,),
27	frozen_stages=1,	27	frozen_stages=1,
28	norm_cfg=norm_cfg,	28	norm_cfg=norm_cfg,
29	norm_eval=True,	29	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/fast_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -23,7 +23,7 @@ model = dict(
23	num_stages=3,	23	num_stages=3,
24	strides=(1, 2, 2),	24	strides=(1, 2, 2),
25	dilations=(1, 1, 1),	25	dilations=(1, 1, 1),
26	- out_indices=(2, ),	26	+ out_indices=(2,),
27	frozen_stages=1,	27	frozen_stages=1,
28	norm_cfg=norm_cfg,	28	norm_cfg=norm_cfg,
29	norm_eval=True,	29	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/faster_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -23,7 +23,7 @@ model = dict(
23	num_stages=3,	23	num_stages=3,
24	strides=(1, 2, 2),	24	strides=(1, 2, 2),
25	dilations=(1, 1, 1),	25	dilations=(1, 1, 1),
26	- out_indices=(2, ),	26	+ out_indices=(2,),
27	frozen_stages=1,	27	frozen_stages=1,
28	norm_cfg=norm_cfg,	28	norm_cfg=norm_cfg,
29	norm_eval=True,	29	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/cascade_mask_rcnn_hrnetv2p_w32_20e.py+2-2

	@@ -24,8 +24,8 @@ model = dict(
24	num_modules=1,	24	num_modules=1,
25	num_branches=1,	25	num_branches=1,
26	block='BOTTLENECK',	26	block='BOTTLENECK',
27	- num_blocks=(4, ),	27	+ num_blocks=(4,),
28	- num_channels=(64, )),	28	+ num_channels=(64,)),
29	stage2=dict(	29	stage2=dict(
30	num_modules=1,	30	num_modules=1,
31	num_branches=2,	31	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/cascade_rcnn_hrnetv2p_w32_20e.py+2-2

	@@ -24,8 +24,8 @@ model = dict(
24	num_modules=1,	24	num_modules=1,
25	num_branches=1,	25	num_branches=1,
26	block='BOTTLENECK',	26	block='BOTTLENECK',
27	- num_blocks=(4, ),	27	+ num_blocks=(4,),
28	- num_channels=(64, )),	28	+ num_channels=(64,)),
29	stage2=dict(	29	stage2=dict(
30	num_modules=1,	30	num_modules=1,
31	num_branches=2,	31	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/faster_rcnn_hrnetv2p_w18_1x.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/faster_rcnn_hrnetv2p_w32_1x.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/faster_rcnn_hrnetv2p_w40_1x.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/fcos_hrnetv2p_w32_gn_1x_4gpu.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/htc_hrnetv2p_w32_20e.py+2-2

	@@ -26,8 +26,8 @@ model = dict(
26	num_modules=1,	26	num_modules=1,
27	num_branches=1,	27	num_branches=1,
28	block='BOTTLENECK',	28	block='BOTTLENECK',
29	- num_blocks=(4, ),	29	+ num_blocks=(4,),
30	- num_channels=(64, )),	30	+ num_channels=(64,)),
31	stage2=dict(	31	stage2=dict(
32	num_modules=1,	32	num_modules=1,
33	num_branches=2,	33	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/mask_rcnn_hrnetv2p_w18_1x.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/hrnet/mask_rcnn_hrnetv2p_w32_1x.py+2-2

	@@ -23,8 +23,8 @@ model = dict(
23	num_modules=1,	23	num_modules=1,
24	num_branches=1,	24	num_branches=1,
25	block='BOTTLENECK',	25	block='BOTTLENECK',
26	- num_blocks=(4, ),	26	+ num_blocks=(4,),
27	- num_channels=(64, )),	27	+ num_channels=(64,)),
28	stage2=dict(	28	stage2=dict(
29	num_modules=1,	29	num_modules=1,
30	num_branches=2,	30	num_branches=2,

MPyTorch/contrib/cv/detection/SOLOv2/configs/libra_rcnn/libra_fast_rcnn_r50_fpn_1x.py+1-1

	@@ -123,7 +123,7 @@ data = dict(
123	type=dataset_type,	123	type=dataset_type,
124	ann_file=data_root + 'annotations/instances_train2017.json',	124	ann_file=data_root + 'annotations/instances_train2017.json',
125	proposal_file=data_root +	125	proposal_file=data_root +
126	- 'libra_proposals/rpn_r50_fpn_1x_train2017.pkl',	126	+ 'libra_proposals/rpn_r50_fpn_1x_train2017.pkl',
127	img_prefix=data_root + 'train2017/',	127	img_prefix=data_root + 'train2017/',
128	pipeline=train_pipeline),	128	pipeline=train_pipeline),
129	val=dict(	129	val=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/mask_rcnn_r50_caffe_c4_1x.py+1-1

	@@ -23,7 +23,7 @@ model = dict(
23	num_stages=3,	23	num_stages=3,
24	strides=(1, 2, 2),	24	strides=(1, 2, 2),
25	dilations=(1, 1, 1),	25	dilations=(1, 1, 1),
26	- out_indices=(2, ),	26	+ out_indices=(2,),
27	frozen_stages=1,	27	frozen_stages=1,
28	norm_cfg=norm_cfg,	28	norm_cfg=norm_cfg,
29	norm_eval=True,	29	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/rpn_r50_caffe_c4_1x.py+1-1

	@@ -22,7 +22,7 @@ model = dict(
22	num_stages=3,	22	num_stages=3,
23	strides=(1, 2, 2),	23	strides=(1, 2, 2),
24	dilations=(1, 1, 1),	24	dilations=(1, 1, 1),
25	- out_indices=(2, ),	25	+ out_indices=(2,),
26	frozen_stages=1,	26	frozen_stages=1,
27	norm_cfg=dict(type='BN', requires_grad=False),	27	norm_cfg=dict(type='BN', requires_grad=False),
28	norm_eval=True,	28	norm_eval=True,

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/decoupled_solo_light_dcn_r50_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch',	25	style='pytorch',
26	dcn=dict(	26	dcn=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/decoupled_solo_light_r50_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/decoupled_solo_r101_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=101,	21	depth=101,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/decoupled_solo_r50_fpn_8gpu_1x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/decoupled_solo_r50_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/solo_r101_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=101,	21	depth=101,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solo/solo_r50_fpn_8gpu_3x.py+1-1

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_light_448_r18_fpn_8gpu_3x.py+10-10

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=18,	21	depth=18,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(
	@@ -51,14 +51,14 @@ model = dict(
51	alpha=0.25,	51	alpha=0.25,
52	loss_weight=1.0)),	52	loss_weight=1.0)),
53	mask_feat_head=dict(	53	mask_feat_head=dict(
54	- type='MaskFeatHead',	54	+ type='MaskFeatHead',
55	- in_channels=256,	55	+ in_channels=256,
56	- out_channels=128,	56	+ out_channels=128,
57	- start_level=0,	57	+ start_level=0,
58	- end_level=3,	58	+ end_level=3,
59	- num_classes=128,	59	+ num_classes=128,
60	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	60	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
61	- )	61	+)
62	# training and testing settings	62	# training and testing settings
63	train_cfg = dict()	63	train_cfg = dict()
64	test_cfg = dict(	64	test_cfg = dict(
	@@ -79,7 +79,7 @@ train_pipeline = [
79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),	79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
80	dict(type='Resize',	80	dict(type='Resize',
81	img_scale=[(768, 512), (768, 480), (768, 448),	81	img_scale=[(768, 512), (768, 480), (768, 448),
82	- (768, 416), (768, 384), (768, 352)],	82	+ (768, 416), (768, 384), (768, 352)],
83	multiscale_mode='value',	83	multiscale_mode='value',
84	keep_ratio=True),	84	keep_ratio=True),
85	dict(type='RandomFlip', flip_ratio=0.5),	85	dict(type='RandomFlip', flip_ratio=0.5),

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_light_448_r34_fpn_8gpu_3x.py+10-10

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=34,	21	depth=34,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(
	@@ -51,14 +51,14 @@ model = dict(
51	alpha=0.25,	51	alpha=0.25,
52	loss_weight=1.0)),	52	loss_weight=1.0)),
53	mask_feat_head=dict(	53	mask_feat_head=dict(
54	- type='MaskFeatHead',	54	+ type='MaskFeatHead',
55	- in_channels=256,	55	+ in_channels=256,
56	- out_channels=128,	56	+ out_channels=128,
57	- start_level=0,	57	+ start_level=0,
58	- end_level=3,	58	+ end_level=3,
59	- num_classes=128,	59	+ num_classes=128,
60	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	60	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
61	- )	61	+)
62	# training and testing settings	62	# training and testing settings
63	train_cfg = dict()	63	train_cfg = dict()
64	test_cfg = dict(	64	test_cfg = dict(
	@@ -79,7 +79,7 @@ train_pipeline = [
79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),	79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
80	dict(type='Resize',	80	dict(type='Resize',
81	img_scale=[(768, 512), (768, 480), (768, 448),	81	img_scale=[(768, 512), (768, 480), (768, 448),
82	- (768, 416), (768, 384), (768, 352)],	82	+ (768, 416), (768, 384), (768, 352)],
83	multiscale_mode='value',	83	multiscale_mode='value',
84	keep_ratio=True),	84	keep_ratio=True),
85	dict(type='RandomFlip', flip_ratio=0.5),	85	dict(type='RandomFlip', flip_ratio=0.5),

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_light_448_r50_fpn_8gpu_3x.py+10-10

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(
	@@ -51,14 +51,14 @@ model = dict(
51	alpha=0.25,	51	alpha=0.25,
52	loss_weight=1.0)),	52	loss_weight=1.0)),
53	mask_feat_head=dict(	53	mask_feat_head=dict(
54	- type='MaskFeatHead',	54	+ type='MaskFeatHead',
55	- in_channels=256,	55	+ in_channels=256,
56	- out_channels=128,	56	+ out_channels=128,
57	- start_level=0,	57	+ start_level=0,
58	- end_level=3,	58	+ end_level=3,
59	- num_classes=128,	59	+ num_classes=128,
60	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	60	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
61	- )	61	+)
62	# training and testing settings	62	# training and testing settings
63	train_cfg = dict()	63	train_cfg = dict()
64	test_cfg = dict(	64	test_cfg = dict(
	@@ -79,7 +79,7 @@ train_pipeline = [
79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),	79	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
80	dict(type='Resize',	80	dict(type='Resize',
81	img_scale=[(768, 512), (768, 480), (768, 448),	81	img_scale=[(768, 512), (768, 480), (768, 448),
82	- (768, 416), (768, 384), (768, 352)],	82	+ (768, 416), (768, 384), (768, 352)],
83	multiscale_mode='value',	83	multiscale_mode='value',
84	keep_ratio=True),	84	keep_ratio=True),
85	dict(type='RandomFlip', flip_ratio=0.5),	85	dict(type='RandomFlip', flip_ratio=0.5),

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_light_512_dcn_r50_fpn_8gpu_3x.py+10-10

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch',	25	style='pytorch',
26	dcn=dict(	26	dcn=dict(
	@@ -58,14 +58,14 @@ model = dict(
58	alpha=0.25,	58	alpha=0.25,
59	loss_weight=1.0)),	59	loss_weight=1.0)),
60	mask_feat_head=dict(	60	mask_feat_head=dict(
61	- type='MaskFeatHead',	61	+ type='MaskFeatHead',
62	- in_channels=256,	62	+ in_channels=256,
63	- out_channels=128,	63	+ out_channels=128,
64	- start_level=0,	64	+ start_level=0,
65	- end_level=3,	65	+ end_level=3,
66	- num_classes=128,	66	+ num_classes=128,
67	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	67	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
68	- )	68	+)
69	# training and testing settings	69	# training and testing settings
70	train_cfg = dict()	70	train_cfg = dict()
71	test_cfg = dict(	71	test_cfg = dict(
	@@ -86,7 +86,7 @@ train_pipeline = [
86	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),	86	dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
87	dict(type='Resize',	87	dict(type='Resize',
88	img_scale=[(852, 512), (852, 480), (852, 448),	88	img_scale=[(852, 512), (852, 480), (852, 448),
89	- (852, 416), (852, 384), (852, 352)],	89	+ (852, 416), (852, 384), (852, 352)],
90	multiscale_mode='value',	90	multiscale_mode='value',
91	keep_ratio=True),	91	keep_ratio=True),
92	dict(type='RandomFlip', flip_ratio=0.5),	92	dict(type='RandomFlip', flip_ratio=0.5),

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_r101_dcn_fpn_8gpu_3x.py+10-10

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=101,	21	depth=101,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch',	25	style='pytorch',
26	dcn=dict(	26	dcn=dict(
	@@ -58,15 +58,15 @@ model = dict(
58	alpha=0.25,	58	alpha=0.25,
59	loss_weight=1.0)),	59	loss_weight=1.0)),
60	mask_feat_head=dict(	60	mask_feat_head=dict(
61	- type='MaskFeatHead',	61	+ type='MaskFeatHead',
62	- in_channels=256,	62	+ in_channels=256,
63	- out_channels=128,	63	+ out_channels=128,
64	- start_level=0,	64	+ start_level=0,
65	- end_level=3,	65	+ end_level=3,
66	- num_classes=256,	66	+ num_classes=256,
67	- conv_cfg=dict(type='DCNv2'),	67	+ conv_cfg=dict(type='DCNv2'),
68	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	68	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
69	- )	69	+)
70	# training and testing settings	70	# training and testing settings
71	train_cfg = dict()	71	train_cfg = dict()
72	test_cfg = dict(	72	test_cfg = dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_r101_fpn_8gpu_3x.py+9-9

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=101,	21	depth=101,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(
	@@ -51,14 +51,14 @@ model = dict(
51	alpha=0.25,	51	alpha=0.25,
52	loss_weight=1.0)),	52	loss_weight=1.0)),
53	mask_feat_head=dict(	53	mask_feat_head=dict(
54	- type='MaskFeatHead',	54	+ type='MaskFeatHead',
55	- in_channels=256,	55	+ in_channels=256,
56	- out_channels=128,	56	+ out_channels=128,
57	- start_level=0,	57	+ start_level=0,
58	- end_level=3,	58	+ end_level=3,
59	- num_classes=256,	59	+ num_classes=256,
60	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	60	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
61	- )	61	+)
62	# training and testing settings	62	# training and testing settings
63	train_cfg = dict()	63	train_cfg = dict()
64	test_cfg = dict(	64	test_cfg = dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_r50_fpn_8gpu_1x.py+10-10

	@@ -23,7 +23,7 @@ model = dict(
23	type='ResNet',	23	type='ResNet',
24	depth=50,	24	depth=50,
25	num_stages=4,	25	num_stages=4,
26	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	26	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
27	frozen_stages=1,	27	frozen_stages=1,
28	style='pytorch'),	28	style='pytorch'),
29	neck=dict(	29	neck=dict(
	@@ -54,14 +54,14 @@ model = dict(
54	alpha=0.25,	54	alpha=0.25,
55	loss_weight=1.0)),	55	loss_weight=1.0)),
56	mask_feat_head=dict(	56	mask_feat_head=dict(
57	- type='MaskFeatHead',	57	+ type='MaskFeatHead',
58	- in_channels=256,	58	+ in_channels=256,
59	- out_channels=128,	59	+ out_channels=128,
60	- start_level=0,	60	+ start_level=0,
61	- end_level=3,	61	+ end_level=3,
62	- num_classes=256,	62	+ num_classes=256,
63	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	63	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
64	- )	64	+)
65	# training and testing settings	65	# training and testing settings
66	train_cfg = dict()	66	train_cfg = dict()
67	test_cfg = dict(	67	test_cfg = dict(
	@@ -83,7 +83,7 @@ train_pipeline = [
83	dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),	83	dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
84	dict(type='RandomFlip', flip_ratio=0.5),	84	dict(type='RandomFlip', flip_ratio=0.5),
85	dict(type='Normalize', **img_norm_cfg),	85	dict(type='Normalize', **img_norm_cfg),
86	- dict(type='Pad', size_divisor=1344), # diff	86	+ dict(type='Pad', size_divisor=1344), # diff
87	dict(type='DefaultFormatBundle'),	87	dict(type='DefaultFormatBundle'),
88	dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),	88	dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
89	]	89	]

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_r50_fpn_8gpu_3x.py+9-9

	@@ -20,7 +20,7 @@ model = dict(
20	type='ResNet',	20	type='ResNet',
21	depth=50,	21	depth=50,
22	num_stages=4,	22	num_stages=4,
23	- out_indices=(0, 1, 2, 3), # C2, C3, C4, C5	23	+ out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
24	frozen_stages=1,	24	frozen_stages=1,
25	style='pytorch'),	25	style='pytorch'),
26	neck=dict(	26	neck=dict(
	@@ -51,14 +51,14 @@ model = dict(
51	alpha=0.25,	51	alpha=0.25,
52	loss_weight=1.0)),	52	loss_weight=1.0)),
53	mask_feat_head=dict(	53	mask_feat_head=dict(
54	- type='MaskFeatHead',	54	+ type='MaskFeatHead',
55	- in_channels=256,	55	+ in_channels=256,
56	- out_channels=128,	56	+ out_channels=128,
57	- start_level=0,	57	+ start_level=0,
58	- end_level=3,	58	+ end_level=3,
59	- num_classes=256,	59	+ num_classes=256,
60	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	60	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
61	- )	61	+)
62	# training and testing settings	62	# training and testing settings
63	train_cfg = dict()	63	train_cfg = dict()
64	test_cfg = dict(	64	test_cfg = dict(

MPyTorch/contrib/cv/detection/SOLOv2/configs/solov2/solov2_x101_dcn_fpn_8gpu_3x.py+9-9

	@@ -60,15 +60,15 @@ model = dict(
60	alpha=0.25,	60	alpha=0.25,
61	loss_weight=1.0)),	61	loss_weight=1.0)),
62	mask_feat_head=dict(	62	mask_feat_head=dict(
63	- type='MaskFeatHead',	63	+ type='MaskFeatHead',
64	- in_channels=256,	64	+ in_channels=256,
65	- out_channels=128,	65	+ out_channels=128,
66	- start_level=0,	66	+ start_level=0,
67	- end_level=3,	67	+ end_level=3,
68	- num_classes=256,	68	+ num_classes=256,
69	- conv_cfg=dict(type='DCNv2'),	69	+ conv_cfg=dict(type='DCNv2'),
70	- norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),	70	+ norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
71	- )	71	+)
72	# training and testing settings	72	# training and testing settings
73	train_cfg = dict()	73	train_cfg = dict()
74	test_cfg = dict(	74	test_cfg = dict(

MPyTorch/contrib/cv/detection/SOLOv2/demo/inference_demo.py+0-1

	@@ -15,7 +15,6 @@
15	from mmdet.apis import init_detector, inference_detector, show_result_pyplot, show_result_ins	15	from mmdet.apis import init_detector, inference_detector, show_result_pyplot, show_result_ins
16	import mmcv	16	import mmcv
17		17
18	-
19	config_file = '../configs/solo/decoupled_solo_r50_fpn_8gpu_3x.py'	18	config_file = '../configs/solo/decoupled_solo_r50_fpn_8gpu_3x.py'
20	# download the checkpoint from model zoo and put it in `checkpoints/`	19	# download the checkpoint from model zoo and put it in `checkpoints/`
21	checkpoint_file = '../checkpoints/DECOUPLED_SOLO_R50_3x.pth'	20	checkpoint_file = '../checkpoints/DECOUPLED_SOLO_R50_3x.pth'

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/docs/conf.py+1-0

	@@ -28,6 +28,7 @@
28	#	28	#
29	import os	29	import os
30	import sys	30	import sys
		31	+
31	sys.path.insert(0, os.path.abspath('..'))	32	sys.path.insert(0, os.path.abspath('..'))
32		33
33	version_file = '../mmcv/version.py'	34	version_file = '../mmcv/version.py'

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/examples/train_cifar10.py+1-1

	@@ -31,7 +31,7 @@ from mmcv import Config
31	from mmcv.runner import DistSamplerSeedHook, Runner	31	from mmcv.runner import DistSamplerSeedHook, Runner
32		32
33		33
34	-def accuracy(output, target, topk=(1, )):	34	+def accuracy(output, target, topk=(1,)):
35	"""Computes the precision@k for the specified values of k"""	35	"""Computes the precision@k for the specified values of k"""
36	with torch.no_grad():	36	with torch.no_grad():
37	maxk = max(topk)	37	maxk = max(topk)

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/cnn/resnet.py+2-2

	@@ -259,7 +259,7 @@ class ResNet(nn.Module):
259	for i, num_blocks in enumerate(stage_blocks):	259	for i, num_blocks in enumerate(stage_blocks):
260	stride = strides[i]	260	stride = strides[i]
261	dilation = dilations[i]	261	dilation = dilations[i]
262	- planes = 64 * 2**i	262	+ planes = 64 * 2 ** i
263	res_layer = make_res_layer(	263	res_layer = make_res_layer(
264	block,	264	block,
265	self.inplanes,	265	self.inplanes,
	@@ -274,7 +274,7 @@ class ResNet(nn.Module):
274	self.add_module(layer_name, res_layer)	274	self.add_module(layer_name, res_layer)
275	self.res_layers.append(layer_name)	275	self.res_layers.append(layer_name)
276		276
277	- self.feat_dim = block.expansion * 64 * 2**(len(stage_blocks) - 1)	277	+ self.feat_dim = block.expansion * 64 * 2 ** (len(stage_blocks) - 1)
278		278
279	def init_weights(self, pretrained=None):	279	def init_weights(self, pretrained=None):
280	if isinstance(pretrained, str):	280	if isinstance(pretrained, str):

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/cnn/vgg.py+1-1

	@@ -108,7 +108,7 @@ class VGG(nn.Module):
108	num_modules = num_blocks * (2 + with_bn) + 1	108	num_modules = num_blocks * (2 + with_bn) + 1
109	end_idx = start_idx + num_modules	109	end_idx = start_idx + num_modules
110	dilation = dilations[i]	110	dilation = dilations[i]
111	- planes = 64 * 2**i if i < 4 else 512	111	+ planes = 64 * 2 ** i if i < 4 else 512
112	vgg_layer = make_vgg_layer(	112	vgg_layer = make_vgg_layer(
113	self.inplanes,	113	self.inplanes,
114	planes,	114	planes,

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/fileio/handlers/base.py+0-1

	@@ -17,7 +17,6 @@ from abc import ABCMeta, abstractmethod
17		17
18		18
19	class BaseFileHandler(object):	19	class BaseFileHandler(object):
20	-
21	__metaclass__ = ABCMeta # python 2 compatibility	20	__metaclass__ = ABCMeta # python 2 compatibility
22		21
23	@abstractmethod	22	@abstractmethod

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/fileio/io.py+0-1

	@@ -119,7 +119,6 @@ def _register_handler(handler, file_formats):
119		119
120		120
121	def register_handler(file_formats, **kwargs):	121	def register_handler(file_formats, **kwargs):
122	-
123	def wrap(cls):	122	def wrap(cls):
124	_register_handler(cls(**kwargs), file_formats)	123	_register_handler(cls(**kwargs), file_formats)
125	return cls	124	return cls

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/image/transforms/colorspace.py+0-1

	@@ -120,7 +120,6 @@ def gray2rgb(img):
120		120
121		121
122	def convert_color_factory(src, dst):	122	def convert_color_factory(src, dst):
123	-
124	code = getattr(cv2, 'COLOR_{}2{}'.format(src.upper(), dst.upper()))	123	code = getattr(cv2, 'COLOR_{}2{}'.format(src.upper(), dst.upper()))
125		124
126	def convert_color(img):	125	def convert_color(img):

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/image/transforms/geometry.py+3-3

	@@ -163,13 +163,13 @@ def imcrop(img, bboxes, scale=1.0, pad_fill=None):
163	patch_shape = (_y2 - _y1 + 1, _x2 - _x1 + 1, chn)	163	patch_shape = (_y2 - _y1 + 1, _x2 - _x1 + 1, chn)
164	patch = np.array(	164	patch = np.array(
165	pad_fill, dtype=img.dtype) * np.ones(	165	pad_fill, dtype=img.dtype) * np.ones(
166	- patch_shape, dtype=img.dtype)	166	+ patch_shape, dtype=img.dtype)
167	x_start = 0 if _x1 >= 0 else -_x1	167	x_start = 0 if _x1 >= 0 else -_x1
168	y_start = 0 if _y1 >= 0 else -_y1	168	y_start = 0 if _y1 >= 0 else -_y1
169	w = x2 - x1 + 1	169	w = x2 - x1 + 1
170	h = y2 - y1 + 1	170	h = y2 - y1 + 1
171	patch[y_start:y_start + h, x_start:x_start + w,	171	patch[y_start:y_start + h, x_start:x_start + w,
172	- ...] = img[y1:y1 + h, x1:x1 + w, ...]	172	+ ...] = img[y1:y1 + h, x1:x1 + w, ...]
173	patches.append(patch)	173	patches.append(patch)
174		174
175	if bboxes.ndim == 1:	175	if bboxes.ndim == 1:
	@@ -192,7 +192,7 @@ def impad(img, shape, pad_val=0):
192	if not isinstance(pad_val, (int, float)):	192	if not isinstance(pad_val, (int, float)):
193	assert len(pad_val) == img.shape[-1]	193	assert len(pad_val) == img.shape[-1]
194	if len(shape) < len(img.shape):	194	if len(shape) < len(img.shape):
195	- shape = shape + (img.shape[-1], )	195	+ shape = shape + (img.shape[-1],)
196	assert len(shape) == len(img.shape)	196	assert len(shape) == len(img.shape)
197	for i in range(len(shape)):	197	for i in range(len(shape)):
198	assert shape[i] >= img.shape[i]	198	assert shape[i] >= img.shape[i]

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/parallel/data_container.py+0-1

	@@ -19,7 +19,6 @@ import torch
19		19
20		20
21	def assert_tensor_type(func):	21	def assert_tensor_type(func):
22	-
23	@functools.wraps(func)	22	@functools.wraps(func)
24	def wrapper(args, *kwargs):	23	def wrapper(args, *kwargs):
25	if not isinstance(args[0].data, torch.Tensor):	24	if not isinstance(args[0].data, torch.Tensor):

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/parallel/data_parallel.py+0-1

	@@ -102,4 +102,3 @@ class MMDataParallel(DataParallel):
102	#	102	#
103	# inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)	103	# inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
104	# return self.module.val_step(inputs[0], *kwargs[0])	104	# return self.module.val_step(inputs[0], *kwargs[0])
105	-

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/parallel/distributed.py+1-1

	@@ -70,4 +70,4 @@ class MMDistributedDataParallel(nn.Module):
70		70
71	# npu_diff	71	# npu_diff
72	inputs, kwargs = self.scatter(inputs, kwargs, [-1])	72	inputs, kwargs = self.scatter(inputs, kwargs, [-1])
73	- return self.module(inputs[0], *kwargs[0])	73	+ return self.module(inputs[0], *kwargs[0])

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/checkpoint.py+42-21

	@@ -29,27 +29,48 @@ import mmcv
29	from .dist_utils import get_dist_info	29	from .dist_utils import get_dist_info
30		30
31	open_mmlab_model_urls = {	31	open_mmlab_model_urls = {
32	- 'vgg16_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/vgg16_caffe-292e1171.pth', # noqa: E501	32	+ 'vgg16_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/vgg16_caffe-292e1171.pth',
33	- 'resnet50_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_caffe-788b5fa3.pth', # noqa: E501	33	+ # noqa: E501
34	- 'resnet101_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_caffe-3ad79236.pth', # noqa: E501	34	+ 'resnet50_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_caffe-788b5fa3.pth',
35	- 'resnext50_32x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50-32x4d-0ab1a123.pth', # noqa: E501	35	+ # noqa: E501
36	- 'resnext101_32x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d-a5af3160.pth', # noqa: E501	36	+ 'resnet101_caffe': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_caffe-3ad79236.pth',
37	- 'resnext101_64x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_64x4d-ee2c6f71.pth', # noqa: E501	37	+ # noqa: E501
38	- 'contrib/resnet50_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn_thangvubk-ad1730dd.pth', # noqa: E501	38	+ 'resnext50_32x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50-32x4d-0ab1a123.pth',
39	- 'detectron/resnet50_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn-9186a21c.pth', # noqa: E501	39	+ # noqa: E501
40	- 'detectron/resnet101_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_gn-cac0ab98.pth', # noqa: E501	40	+ 'resnext101_32x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d-a5af3160.pth',
41	- 'jhu/resnet50_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn_ws-15beedd8.pth', # noqa: E501	41	+ # noqa: E501
42	- 'jhu/resnet101_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_gn_ws-3e3c308c.pth', # noqa: E501	42	+ 'resnext101_64x4d': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_64x4d-ee2c6f71.pth',
43	- 'jhu/resnext50_32x4d_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50_32x4d_gn_ws-0d87ac85.pth', # noqa: E501	43	+ # noqa: E501
44	- 'jhu/resnext101_32x4d_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d_gn_ws-34ac1a9e.pth', # noqa: E501	44	+ 'contrib/resnet50_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn_thangvubk-ad1730dd.pth',
45	- 'jhu/resnext50_32x4d_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50_32x4d_gn-c7e8b754.pth', # noqa: E501	45	+ # noqa: E501
46	- 'jhu/resnext101_32x4d_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d_gn-ac3bb84e.pth', # noqa: E501	46	+ 'detectron/resnet50_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn-9186a21c.pth',
47	- 'msra/hrnetv2_w18': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w18-00eb2006.pth', # noqa: E501	47	+ # noqa: E501
48	- 'msra/hrnetv2_w32': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w32-dc9eeb4f.pth', # noqa: E501	48	+ 'detectron/resnet101_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_gn-cac0ab98.pth',
49	- 'msra/hrnetv2_w40': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w40-ed0b031c.pth', # noqa: E501	49	+ # noqa: E501
50	- 'bninception_caffe': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/bn_inception_caffe-ed2e8665.pth', # noqa: E501	50	+ 'jhu/resnet50_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet50_gn_ws-15beedd8.pth',
51	- 'kin400/i3d_r50_f32s2_k400': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/i3d_r50_f32s2_k400-2c57e077.pth', # noqa: E501	51	+ # noqa: E501
52	- 'kin400/nl3d_r50_f32s2_k400': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/nl3d_r50_f32s2_k400-fa7e7caa.pth', # noqa: E501	52	+ 'jhu/resnet101_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnet101_gn_ws-3e3c308c.pth',
		53	+ # noqa: E501
		54	+ 'jhu/resnext50_32x4d_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50_32x4d_gn_ws-0d87ac85.pth',
		55	+ # noqa: E501
		56	+ 'jhu/resnext101_32x4d_gn_ws': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d_gn_ws-34ac1a9e.pth',
		57	+ # noqa: E501
		58	+ 'jhu/resnext50_32x4d_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext50_32x4d_gn-c7e8b754.pth',
		59	+ # noqa: E501
		60	+ 'jhu/resnext101_32x4d_gn': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d_gn-ac3bb84e.pth',
		61	+ # noqa: E501
		62	+ 'msra/hrnetv2_w18': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w18-00eb2006.pth',
		63	+ # noqa: E501
		64	+ 'msra/hrnetv2_w32': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w32-dc9eeb4f.pth',
		65	+ # noqa: E501
		66	+ 'msra/hrnetv2_w40': 'https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/hrnetv2_w40-ed0b031c.pth',
		67	+ # noqa: E501
		68	+ 'bninception_caffe': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/bn_inception_caffe-ed2e8665.pth',
		69	+ # noqa: E501
		70	+ 'kin400/i3d_r50_f32s2_k400': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/i3d_r50_f32s2_k400-2c57e077.pth',
		71	+ # noqa: E501
		72	+ 'kin400/nl3d_r50_f32s2_k400': 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/nl3d_r50_f32s2_k400-fa7e7caa.pth',
		73	+ # noqa: E501
53	} # yapf: disable	74	} # yapf: disable
54		75
55		76

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/dist_utils.py+3-1

	@@ -18,6 +18,9 @@ import os
18	import subprocess	18	import subprocess
19		19
20	import torch	20	import torch
		21	+
		22	+if torch.__version__ >= '1.8.1':
		23	+ import torch_npu
21	import torch.distributed as dist	24	import torch.distributed as dist
22	import torch.multiprocessing as mp	25	import torch.multiprocessing as mp
23		26
	@@ -87,7 +90,6 @@ def get_dist_info():
87		90
88		91
89	def master_only(func):	92	def master_only(func):
90	-
91	@functools.wraps(func)	93	@functools.wraps(func)
92	def wrapper(args, *kwargs):	94	def wrapper(args, *kwargs):
93	rank, _ = get_dist_info()	95	rank, _ = get_dist_info()

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/hooks/iter_timer.py+2-1

	@@ -24,12 +24,13 @@ class IterTimerHook(Hook):
24	self.t = time.time()	24	self.t = time.time()
25	self.skip_step = 0	25	self.skip_step = 0
26	self.time_all = 0	26	self.time_all = 0
		27	+
27	def before_iter(self, runner):	28	def before_iter(self, runner):
28	runner.log_buffer.update({'data_time': time.time() - self.t})	29	runner.log_buffer.update({'data_time': time.time() - self.t})
29		30
30	def after_iter(self, runner):	31	def after_iter(self, runner):
31	## npu diff	32	## npu diff
32	- #runner.log_buffer.update({'time': time.time() - self.t})	33	+ # runner.log_buffer.update({'time': time.time() - self.t})
33	cur_time = time.time()	34	cur_time = time.time()
34	runner.log_buffer.update({'time': time.time() - self.t})	35	runner.log_buffer.update({'time': time.time() - self.t})
35	if self.skip_step >= 5:	36	if self.skip_step >= 5:

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/hooks/logger/base.py+1-1

	@@ -33,7 +33,7 @@ class LoggerHook(Hook):
33	def __init__(self, interval=10, ignore_last=True, reset_flag=False):	33	def __init__(self, interval=10, ignore_last=True, reset_flag=False):
34	self.interval = interval	34	self.interval = interval
35	self.ignore_last = ignore_last	35	self.ignore_last = ignore_last
36	- self.reset_flag = reset_flag	36	+ self.reset_flag = False # reset_flag
37		37
38	@abstractmethod	38	@abstractmethod
39	def log(self, runner):	39	def log(self, runner):

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/hooks/logger/pavi.py+2-2

	@@ -47,7 +47,7 @@ class PaviClient(object):
47	if not var:	47	if not var:
48	raise ValueError(	48	raise ValueError(
49	'"{}" is neither specified nor defined as env variables'.	49	'"{}" is neither specified nor defined as env variables'.
50	- format(env_var))	50	+ format(env_var))
51	return var	51	return var
52		52
53	def _print_log(self, msg, level=logging.INFO, args, *kwargs):	53	def _print_log(self, msg, level=logging.INFO, args, *kwargs):
	@@ -129,7 +129,7 @@ class PaviClient(object):
129	else:	129	else:
130	self._print_log(	130	self._print_log(
131	'unexpected status code: {}, err msg: {}'.	131	'unexpected status code: {}, err msg: {}'.
132	- format(status_code, response.reason),	132	+ format(status_code, response.reason),
133	level=logging.ERROR)	133	level=logging.ERROR)
134	retry += 1	134	retry += 1
135	if retry == max_retry:	135	if retry == max_retry:

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/hooks/logger/text.py+3-3

	@@ -53,7 +53,7 @@ class TextLoggerHook(LoggerHook):
53	if 'time' in log_dict.keys():	53	if 'time' in log_dict.keys():
54	self.time_sec_tot += (log_dict['time'] * self.interval)	54	self.time_sec_tot += (log_dict['time'] * self.interval)
55	time_sec_avg = self.time_sec_tot / (	55	time_sec_avg = self.time_sec_tot / (
56	- runner.iter - self.start_iter + 1)	56	+ runner.iter - self.start_iter + 1)
57	eta_sec = time_sec_avg * (runner.max_iters - runner.iter - 1)	57	eta_sec = time_sec_avg * (runner.max_iters - runner.iter - 1)
58	eta_str = str(datetime.timedelta(seconds=int(eta_sec)))	58	eta_str = str(datetime.timedelta(seconds=int(eta_sec)))
59	log_str += 'eta: {}, '.format(eta_str)	59	log_str += 'eta: {}, '.format(eta_str)
	@@ -71,8 +71,8 @@ class TextLoggerHook(LoggerHook):
71	# TODO: resolve this hack	71	# TODO: resolve this hack
72	# these items have been in log_str	72	# these items have been in log_str
73	if name in [	73	if name in [
74	- 'mode', 'Epoch', 'iter', 'lr', 'time', 'data_time',	74	+ 'mode', 'Epoch', 'iter', 'lr', 'time', 'data_time',
75	- 'memory', 'epoch'	75	+ 'memory', 'epoch'
76	]:	76	]:
77	continue	77	continue
78	if isinstance(val, float):	78	if isinstance(val, float):

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/hooks/lr_updater.py+7-7

	@@ -64,7 +64,7 @@ class LrUpdaterHook(Hook):
64	k = (1 - cur_iters / self.warmup_iters) * (1 - self.warmup_ratio)	64	k = (1 - cur_iters / self.warmup_iters) * (1 - self.warmup_ratio)
65	warmup_lr = [_lr * (1 - k) for _lr in self.regular_lr]	65	warmup_lr = [_lr * (1 - k) for _lr in self.regular_lr]
66	elif self.warmup == 'exp':	66	elif self.warmup == 'exp':
67	- k = self.warmup_ratio**(1 - cur_iters / self.warmup_iters)	67	+ k = self.warmup_ratio ** (1 - cur_iters / self.warmup_iters)
68	warmup_lr = [_lr * k for _lr in self.regular_lr]	68	warmup_lr = [_lr * k for _lr in self.regular_lr]
69	return warmup_lr	69	return warmup_lr
70		70
	@@ -130,14 +130,14 @@ class StepLrUpdaterHook(LrUpdaterHook):
130	progress = runner.epoch if self.by_epoch else runner.iter	130	progress = runner.epoch if self.by_epoch else runner.iter
131		131
132	if isinstance(self.step, int):	132	if isinstance(self.step, int):
133	- return base_lr * (self.gamma**(progress // self.step))	133	+ return base_lr * (self.gamma ** (progress // self.step))
134		134
135	exp = len(self.step)	135	exp = len(self.step)
136	for i, s in enumerate(self.step):	136	for i, s in enumerate(self.step):
137	if progress < s:	137	if progress < s:
138	exp = i	138	exp = i
139	break	139	break
140	- return base_lr * self.gamma**exp	140	+ return base_lr * self.gamma ** exp
141		141
142		142
143	class ExpLrUpdaterHook(LrUpdaterHook):	143	class ExpLrUpdaterHook(LrUpdaterHook):
	@@ -148,7 +148,7 @@ class ExpLrUpdaterHook(LrUpdaterHook):
148		148
149	def get_lr(self, runner, base_lr):	149	def get_lr(self, runner, base_lr):
150	progress = runner.epoch if self.by_epoch else runner.iter	150	progress = runner.epoch if self.by_epoch else runner.iter
151	- return base_lr * self.gamma**progress	151	+ return base_lr * self.gamma ** progress
152		152
153		153
154	class PolyLrUpdaterHook(LrUpdaterHook):	154	class PolyLrUpdaterHook(LrUpdaterHook):
	@@ -165,7 +165,7 @@ class PolyLrUpdaterHook(LrUpdaterHook):
165	else:	165	else:
166	progress = runner.iter	166	progress = runner.iter
167	max_progress = runner.max_iters	167	max_progress = runner.max_iters
168	- coeff = (1 - progress / max_progress)**self.power	168	+ coeff = (1 - progress / max_progress) ** self.power
169	return (base_lr - self.min_lr) * coeff + self.min_lr	169	return (base_lr - self.min_lr) * coeff + self.min_lr
170		170
171		171
	@@ -178,7 +178,7 @@ class InvLrUpdaterHook(LrUpdaterHook):
178		178
179	def get_lr(self, runner, base_lr):	179	def get_lr(self, runner, base_lr):
180	progress = runner.epoch if self.by_epoch else runner.iter	180	progress = runner.epoch if self.by_epoch else runner.iter
181	- return base_lr * (1 + self.gamma * progress)**(-self.power)	181	+ return base_lr * (1 + self.gamma * progress) ** (-self.power)
182		182
183		183
184	class CosineLrUpdaterHook(LrUpdaterHook):	184	class CosineLrUpdaterHook(LrUpdaterHook):
	@@ -195,4 +195,4 @@ class CosineLrUpdaterHook(LrUpdaterHook):
195	progress = runner.iter	195	progress = runner.iter
196	max_progress = runner.max_iters	196	max_progress = runner.max_iters
197	return self.target_lr + 0.5 * (base_lr - self.target_lr) * \	197	return self.target_lr + 0.5 * (base_lr - self.target_lr) * \
198	- (1 + cos(pi * (progress / max_progress)))	198	+ (1 + cos(pi * (progress / max_progress)))

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/runner/runner.py+16-8

	@@ -81,10 +81,10 @@ class Runner(object):
81		81
82	self._rank, self._world_size = get_dist_info()	82	self._rank, self._world_size = get_dist_info()
83	self.timestamp = get_time_str()	83	self.timestamp = get_time_str()
84	- if logger is None:	84	+ # if logger is None:
85	- self.logger = self.init_logger(work_dir, log_level)	85	+ self.logger = self.init_logger(work_dir, log_level)
86	- else:	86	+ # else:
87	- self.logger = logger	87	+ # self.logger = logger
88	self.log_buffer = LogBuffer()	88	self.log_buffer = LogBuffer()
89		89
90	self.mode = None	90	self.mode = None
	@@ -94,9 +94,11 @@ class Runner(object):
94	self._inner_iter = 0	94	self._inner_iter = 0
95	self._max_epochs = 0	95	self._max_epochs = 0
96	self._max_iters = 0	96	self._max_iters = 0
		97	+ self.train_performance = False
97	self.samples_per_gpu = samples_per_gpu	98	self.samples_per_gpu = samples_per_gpu
98	self.num_of_gpus = num_of_gpus	99	self.num_of_gpus = num_of_gpus
99	self.iter_time_hook = IterTimerHook()	100	self.iter_time_hook = IterTimerHook()
		101	+
100	@property	102	@property
101	def model_name(self):	103	def model_name(self):
102	"""str: Name of the model, usually the module class name."""	104	"""str: Name of the model, usually the module class name."""
	@@ -194,6 +196,9 @@ class Runner(object):
194	logging.basicConfig(	196	logging.basicConfig(
195	format='%(asctime)s - %(levelname)s - %(message)s', level=level)	197	format='%(asctime)s - %(levelname)s - %(message)s', level=level)
196	logger = logging.getLogger(__name__)	198	logger = logging.getLogger(__name__)
		199	+ if self.rank == 0:
		200	+ logger.addHandler(logging.StreamHandler())
		201	+ logger.setLevel(logging.INFO)
197	if log_dir and self.rank == 0:	202	if log_dir and self.rank == 0:
198	filename = '{}.log'.format(self.timestamp)	203	filename = '{}.log'.format(self.timestamp)
199	log_file = osp.join(log_dir, filename)	204	log_file = osp.join(log_dir, filename)
	@@ -325,11 +330,14 @@ class Runner(object):
325	self.outputs = outputs	330	self.outputs = outputs
326	self.call_hook('after_train_iter')	331	self.call_hook('after_train_iter')
327	self._iter += 1	332	self._iter += 1
		333	+ if i >= 500 and self.train_performance:
		334	+ exit(0)
328	if i % 200 == 0 and i:	335	if i % 200 == 0 and i:
329	self.logger.info('FPS: %02f' % (self.samples_per_gpu * self.num_of_gpus * (i - 5) /	336	self.logger.info('FPS: %02f' % (self.samples_per_gpu * self.num_of_gpus * (i - 5) /
330	- self.iter_time_hook.time_all))	337	+ self.iter_time_hook.time_all))
331		338
332	- self.logger.info('FPS: ' + str(self.samples_per_gpu * self.num_of_gpus / self.iter_time_hook.time_all * (len(self.data_loader) - 5)))	339	+ self.logger.info('FPS: ' + str(
		340	+ self.samples_per_gpu * self.num_of_gpus / self.iter_time_hook.time_all * (len(self.data_loader) - 5)))
333	self.call_hook('after_train_epoch')	341	self.call_hook('after_train_epoch')
334	self._epoch += 1	342	self._epoch += 1
335		343
	@@ -405,14 +413,14 @@ class Runner(object):
405	if not hasattr(self, mode):	413	if not hasattr(self, mode):
406	raise ValueError(	414	raise ValueError(
407	'runner has no method named "{}" to run an epoch'.	415	'runner has no method named "{}" to run an epoch'.
408	- format(mode))	416	+ format(mode))
409	epoch_runner = getattr(self, mode)	417	epoch_runner = getattr(self, mode)
410	elif callable(mode): # custom train()	418	elif callable(mode): # custom train()
411	epoch_runner = mode	419	epoch_runner = mode
412	else:	420	else:
413	raise TypeError('mode in workflow must be a str or '	421	raise TypeError('mode in workflow must be a str or '
414	'callable function, not {}'.format(	422	'callable function, not {}'.format(
415	- type(mode)))	423	+ type(mode)))
416	for _ in range(epochs):	424	for _ in range(epochs):
417	if mode == 'train' and self.epoch >= max_epochs:	425	if mode == 'train' and self.epoch >= max_epochs:
418	return	426	return

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/utils/misc.py+1-1

	@@ -158,7 +158,7 @@ def check_prerequisites(
158	prerequisites,	158	prerequisites,
159	checker,	159	checker,
160	msg_tmpl='Prerequisites "{}" are required in method "{}" but not '	160	msg_tmpl='Prerequisites "{}" are required in method "{}" but not '
161	- 'found, please install them first.'): # yapf: disable	161	+ 'found, please install them first.'): # yapf: disable
162	"""A decorator factory to check if prerequisites are satisfied.	162	"""A decorator factory to check if prerequisites are satisfied.
163		163
164	Args:	164	Args:

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/utils/path.py+1-1

	@@ -94,7 +94,7 @@ def scandir(dir_path, suffix=None):
94	return _scandir_py(dir_path, suffix)	94	return _scandir_py(dir_path, suffix)
95		95
96		96
97	-def find_vcs_root(path, markers=('.git', )):	97	+def find_vcs_root(path, markers=('.git',)):
98	"""Finds the root directory (including itself) of specified markers.	98	"""Finds the root directory (including itself) of specified markers.
99		99
100	Args:	100	Args:

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/video/__init__.py+0-10

	@@ -13,13 +13,3 @@
13	# limitations under the License.	13	# limitations under the License.
14		14
15	# Copyright (c) Open-MMLab. All rights reserved.	15	# Copyright (c) Open-MMLab. All rights reserved.
16	-from .io import Cache, VideoReader, frames2video
17	-from .optflow import (dequantize_flow, flow_warp, flowread, flowwrite,
18	- quantize_flow)
19	-from .processing import concat_video, convert_video, cut_video, resize_video
20	-
21	-__all__ = [
22	- 'Cache', 'VideoReader', 'frames2video', 'convert_video', 'resize_video',
23	- 'cut_video', 'concat_video', 'flowread', 'flowwrite', 'quantize_flow',
24	- 'dequantize_flow', 'flow_warp'
25	-]

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/video/optflow.py+2-2

	@@ -54,7 +54,7 @@ def flowread(flow_or_path, quantize=False, concat_axis=0, args, *kwargs):
54	if header != 'PIEH':	54	if header != 'PIEH':
55	raise IOError(	55	raise IOError(
56	'Invalid flow file: {}, header does not contain PIEH'.	56	'Invalid flow file: {}, header does not contain PIEH'.
57	- format(flow_or_path))	57	+ format(flow_or_path))
58		58
59	w = np.fromfile(f, np.int32, 1).squeeze()	59	w = np.fromfile(f, np.int32, 1).squeeze()
60	h = np.fromfile(f, np.int32, 1).squeeze()	60	h = np.fromfile(f, np.int32, 1).squeeze()
	@@ -65,7 +65,7 @@ def flowread(flow_or_path, quantize=False, concat_axis=0, args, *kwargs):
65	if cat_flow.ndim != 2:	65	if cat_flow.ndim != 2:
66	raise IOError(	66	raise IOError(
67	'{} is not a valid quantized flow file, its dimension is {}.'.	67	'{} is not a valid quantized flow file, its dimension is {}.'.
68	- format(flow_or_path, cat_flow.ndim))	68	+ format(flow_or_path, cat_flow.ndim))
69	assert cat_flow.shape[concat_axis] % 2 == 0	69	assert cat_flow.shape[concat_axis] % 2 == 0
70	dx, dy = np.split(cat_flow, 2, axis=concat_axis)	70	dx, dy = np.split(cat_flow, 2, axis=concat_axis)
71	flow = dequantize_flow(dx, dy, args, *kwargs)	71	flow = dequantize_flow(dx, dy, args, *kwargs)

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/mmcv/visualization/optflow.py+5-5

	@@ -18,7 +18,7 @@ from __future__ import division
18	import numpy as np	18	import numpy as np
19		19
20	from mmcv.image import rgb2bgr	20	from mmcv.image import rgb2bgr
21	-from mmcv.video import flowread	21	+# from mmcv.video import flowread
22	from .image import imshow	22	from .image import imshow
23		23
24		24
	@@ -58,12 +58,12 @@ def flow2rgb(flow, color_wheel=None, unknown_thr=1e6):
58	dy = flow[:, :, 1].copy()	58	dy = flow[:, :, 1].copy()
59		59
60	ignore_inds = (	60	ignore_inds = (
61	- np.isnan(dx) \| np.isnan(dy) \| (np.abs(dx) > unknown_thr) \|	61	+ np.isnan(dx) \| np.isnan(dy) \| (np.abs(dx) > unknown_thr) \|
62	- (np.abs(dy) > unknown_thr))	62	+ (np.abs(dy) > unknown_thr))
63	dx[ignore_inds] = 0	63	dx[ignore_inds] = 0
64	dy[ignore_inds] = 0	64	dy[ignore_inds] = 0
65		65
66	- rad = np.sqrt(dx2 + dy2)	66	+ rad = np.sqrt(dx 2 + dy 2)
67	if np.any(rad > np.finfo(float).eps):	67	if np.any(rad > np.finfo(float).eps):
68	max_rad = np.max(rad)	68	max_rad = np.max(rad)
69	dx /= max_rad	69	dx /= max_rad
	@@ -71,7 +71,7 @@ def flow2rgb(flow, color_wheel=None, unknown_thr=1e6):
71		71
72	[h, w] = dx.shape	72	[h, w] = dx.shape
73		73
74	- rad = np.sqrt(dx2 + dy2)	74	+ rad = np.sqrt(dx 2 + dy 2)
75	angle = np.arctan2(-dy, -dx) / np.pi	75	angle = np.arctan2(-dy, -dx) / np.pi
76		76
77	bin_real = (angle + 1) / 2 * (num_bins - 1)	77	bin_real = (angle + 1) / 2 * (num_bins - 1)

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/setup.py+0-11

	@@ -79,17 +79,6 @@ else:
79	extra_link_args = []	79	extra_link_args = []
80		80
81	EXT_MODULES = [	81	EXT_MODULES = [
82	- Extension(
83	- name='mmcv._ext',
84	- sources=[
85	- './mmcv/video/optflow_warp/flow_warp.cpp',
86	- './mmcv/video/optflow_warp/flow_warp_module.pyx'
87	- ],
88	- include_dirs=[numpy.get_include()],
89	- language='c++',
90	- extra_compile_args=extra_compile_args,
91	- extra_link_args=extra_link_args,
92	- ),
93	]	82	]
94		83
95	setup(	84	setup(

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/tests/data/config/a.b.py+0-1

	@@ -11,4 +11,3 @@
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.	11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and	12	# See the License for the specific language governing permissions and
13	# limitations under the License.	13	# limitations under the License.
14	-

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/tests/test_fileio.py+0-4

	@@ -59,7 +59,6 @@ obj_for_test = [{'a': 'abc', 'b': 1}, 2, 'c']
59		59
60		60
61	def test_json():	61	def test_json():
62	-
63	def json_checker(dump_str):	62	def json_checker(dump_str):
64	assert dump_str in [	63	assert dump_str in [
65	'[{"a": "abc", "b": 1}, 2, "c"]', '[{"b": 1, "a": "abc"}, 2, "c"]'	64	'[{"a": "abc", "b": 1}, 2, "c"]', '[{"b": 1, "a": "abc"}, 2, "c"]'
	@@ -69,7 +68,6 @@ def test_json():
69		68
70		69
71	def test_yaml():	70	def test_yaml():
72	-
73	def yaml_checker(dump_str):	71	def yaml_checker(dump_str):
74	assert dump_str in [	72	assert dump_str in [
75	'- {a: abc, b: 1}\n- 2\n- c\n', '- {b: 1, a: abc}\n- 2\n- c\n',	73	'- {a: abc, b: 1}\n- 2\n- c\n', '- {b: 1, a: abc}\n- 2\n- c\n',
	@@ -80,7 +78,6 @@ def test_yaml():
80		78
81		79
82	def test_pickle():	80	def test_pickle():
83	-
84	def pickle_checker(dump_str):	81	def pickle_checker(dump_str):
85	import pickle	82	import pickle
86	assert pickle.loads(dump_str) == obj_for_test	83	assert pickle.loads(dump_str) == obj_for_test
	@@ -99,7 +96,6 @@ def test_exception():
99		96
100		97
101	def test_register_handler():	98	def test_register_handler():
102	-
103	@mmcv.register_handler('txt')	99	@mmcv.register_handler('txt')
104	class TxtHandler1(mmcv.BaseFileHandler):	100	class TxtHandler1(mmcv.BaseFileHandler):
105		101

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/tests/test_image.py+5-5

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/tests/test_misc.py+1-3

	@@ -32,7 +32,7 @@ def test_iter_cast():
32		32
33	def test_is_seq_of():	33	def test_is_seq_of():
34	assert mmcv.is_seq_of([1.0, 2.0, 3.0], float)	34	assert mmcv.is_seq_of([1.0, 2.0, 3.0], float)
35	- assert mmcv.is_seq_of([(1, ), (2, ), (3, )], tuple)	35	+ assert mmcv.is_seq_of([(1,), (2,), (3,)], tuple)
36	assert mmcv.is_seq_of((1.0, 2.0, 3.0), float)	36	assert mmcv.is_seq_of((1.0, 2.0, 3.0), float)
37	assert mmcv.is_list_of([1.0, 2.0, 3.0], float)	37	assert mmcv.is_list_of([1.0, 2.0, 3.0], float)
38	assert not mmcv.is_seq_of((1.0, 2.0, 3.0), float, seq_type=list)	38	assert not mmcv.is_seq_of((1.0, 2.0, 3.0), float, seq_type=list)
	@@ -57,7 +57,6 @@ def test_concat_list():
57		57
58		58
59	def test_requires_package(capsys):	59	def test_requires_package(capsys):
60	-
61	@mmcv.requires_package('nnn')	60	@mmcv.requires_package('nnn')
62	def func_a():	61	def func_a():
63	pass	62	pass
	@@ -87,7 +86,6 @@ def test_requires_package(capsys):
87		86
88		87
89	def test_requires_executable(capsys):	88	def test_requires_executable(capsys):
90	-
91	@mmcv.requires_executable('nnn')	89	@mmcv.requires_executable('nnn')
92	def func_a():	90	def func_a():
93	pass	91	pass

MPyTorch/contrib/cv/detection/SOLOv2/mmcv/tests/test_optflow.py+68-69

	@@ -158,7 +158,6 @@ def test_flow2rgb():
158		158
159		159
160	def test_flow_warp():	160	def test_flow_warp():
161	-
162	def np_flow_warp(flow, img):	161	def np_flow_warp(flow, img):
163	output = np.zeros_like(img, dtype=img.dtype)	162	output = np.zeros_like(img, dtype=img.dtype)
164	height = flow.shape[0]	163	height = flow.shape[0]
	@@ -172,7 +171,7 @@ def test_flow_warp():
172	valid = (sx >= 0) & (sx < height - 1) & (sy >= 0) & (sy < width - 1)	171	valid = (sx >= 0) & (sx < height - 1) & (sy >= 0) & (sy < width - 1)
173		172
174	output[valid, :] = img[dx[valid].round().astype(int),	173	output[valid, :] = img[dx[valid].round().astype(int),
175	- dy[valid].round().astype(int), :]	174	+ dy[valid].round().astype(int), :]
176		175
177	return output	176	return output
178		177
	@@ -202,74 +201,74 @@ def test_make_color_wheel():
202	color_wheel = mmcv.make_color_wheel([2, 2, 2, 2, 2, 2])	201	color_wheel = mmcv.make_color_wheel([2, 2, 2, 2, 2, 2])
203	# yapf: disable	202	# yapf: disable
204	assert_array_equal(default_color_wheel, np.array(	203	assert_array_equal(default_color_wheel, np.array(
205	- [[1. , 0. , 0. ],	204	+ [[1., 0., 0.],
206	- [1. , 0.06666667, 0. ],	205	+ [1., 0.06666667, 0.],
207	- [1. , 0.13333334, 0. ],	206	+ [1., 0.13333334, 0.],
208	- [1. , 0.2 , 0. ],	207	+ [1., 0.2, 0.],
209	- [1. , 0.26666668, 0. ],	208	+ [1., 0.26666668, 0.],
210	- [1. , 0.33333334, 0. ],	209	+ [1., 0.33333334, 0.],
211	- [1. , 0.4 , 0. ],	210	+ [1., 0.4, 0.],
212	- [1. , 0.46666667, 0. ],	211	+ [1., 0.46666667, 0.],
213	- [1. , 0.53333336, 0. ],	212	+ [1., 0.53333336, 0.],
214	- [1. , 0.6 , 0. ],	213	+ [1., 0.6, 0.],
215	- [1. , 0.6666667 , 0. ],	214	+ [1., 0.6666667, 0.],
216	- [1. , 0.73333335, 0. ],	215	+ [1., 0.73333335, 0.],
217	- [1. , 0.8 , 0. ],	216	+ [1., 0.8, 0.],
218	- [1. , 0.8666667 , 0. ],	217	+ [1., 0.8666667, 0.],
219	- [1. , 0.93333334, 0. ],	218	+ [1., 0.93333334, 0.],
220	- [1. , 1. , 0. ],	219	+ [1., 1., 0.],
221	- [0.8333333 , 1. , 0. ],	220	+ [0.8333333, 1., 0.],
222	- [0.6666667 , 1. , 0. ],	221	+ [0.6666667, 1., 0.],
223	- [0.5 , 1. , 0. ],	222	+ [0.5, 1., 0.],
224	- [0.33333334, 1. , 0. ],	223	+ [0.33333334, 1., 0.],
225	- [0.16666667, 1. , 0. ],	224	+ [0.16666667, 1., 0.],
226	- [0. , 1. , 0. ],	225	+ [0., 1., 0.],
227	- [0. , 1. , 0.25 ],	226	+ [0., 1., 0.25],
228	- [0. , 1. , 0.5 ],	227	+ [0., 1., 0.5],
229	- [0. , 1. , 0.75 ],	228	+ [0., 1., 0.75],
230	- [0. , 1. , 1. ],	229	+ [0., 1., 1.],
231	- [0. , 0.90909094, 1. ],	230	+ [0., 0.90909094, 1.],
232	- [0. , 0.8181818 , 1. ],	231	+ [0., 0.8181818, 1.],
233	- [0. , 0.72727275, 1. ],	232	+ [0., 0.72727275, 1.],
234	- [0. , 0.6363636 , 1. ],	233	+ [0., 0.6363636, 1.],
235	- [0. , 0.54545456, 1. ],	234	+ [0., 0.54545456, 1.],
236	- [0. , 0.45454547, 1. ],	235	+ [0., 0.45454547, 1.],
237	- [0. , 0.36363637, 1. ],	236	+ [0., 0.36363637, 1.],
238	- [0. , 0.27272728, 1. ],	237	+ [0., 0.27272728, 1.],
239	- [0. , 0.18181819, 1. ],	238	+ [0., 0.18181819, 1.],
240	- [0. , 0.09090909, 1. ],	239	+ [0., 0.09090909, 1.],
241	- [0. , 0. , 1. ],	240	+ [0., 0., 1.],
242	- [0.07692308, 0. , 1. ],	241	+ [0.07692308, 0., 1.],
243	- [0.15384616, 0. , 1. ],	242	+ [0.15384616, 0., 1.],
244	- [0.23076923, 0. , 1. ],	243	+ [0.23076923, 0., 1.],
245	- [0.30769232, 0. , 1. ],	244	+ [0.30769232, 0., 1.],
246	- [0.3846154 , 0. , 1. ],	245	+ [0.3846154, 0., 1.],
247	- [0.46153846, 0. , 1. ],	246	+ [0.46153846, 0., 1.],
248	- [0.53846157, 0. , 1. ],	247	+ [0.53846157, 0., 1.],
249	- [0.61538464, 0. , 1. ],	248	+ [0.61538464, 0., 1.],
250	- [0.6923077 , 0. , 1. ],	249	+ [0.6923077, 0., 1.],
251	- [0.7692308 , 0. , 1. ],	250	+ [0.7692308, 0., 1.],
252	- [0.84615386, 0. , 1. ],	251	+ [0.84615386, 0., 1.],
253	- [0.9230769 , 0. , 1. ],	252	+ [0.9230769, 0., 1.],
254	- [1. , 0. , 1. ],	253	+ [1., 0., 1.],
255	- [1. , 0. , 0.8333333 ],	254	+ [1., 0., 0.8333333],
256	- [1. , 0. , 0.6666667 ],	255	+ [1., 0., 0.6666667],
257	- [1. , 0. , 0.5 ],	256	+ [1., 0., 0.5],
258	- [1. , 0. , 0.33333334],	257	+ [1., 0., 0.33333334],
259	- [1. , 0. , 0.16666667]], dtype=np.float32))	258	+ [1., 0., 0.16666667]], dtype=np.float32))
260		259
261	assert_array_equal(	260	assert_array_equal(
262	color_wheel,	261	color_wheel,
263	- np.array([[1., 0. , 0. ],	262	+ np.array([[1., 0., 0.],
264	- [1. , 0.5, 0. ],	263	+ [1., 0.5, 0.],
265	- [1. , 1. , 0. ],	264	+ [1., 1., 0.],
266	- [0.5, 1. , 0. ],	265	+ [0.5, 1., 0.],
267	- [0. , 1. , 0. ],	266	+ [0., 1., 0.],
268	- [0. , 1. , 0.5],	267	+ [0., 1., 0.5],
269	- [0. , 1. , 1. ],	268	+ [0., 1., 1.],
270	- [0. , 0.5, 1. ],	269	+ [0., 0.5, 1.],
271	- [0. , 0. , 1. ],	270	+ [0., 0., 1.],
272	- [0.5, 0. , 1. ],	271	+ [0.5, 0., 1.],
273	- [1. , 0. , 1. ],	272	+ [1., 0., 1.],
274	- [1. , 0. , 0.5]], dtype=np.float32))	273	+ [1., 0., 0.5]], dtype=np.float32))
275	# yapf: enable	274	# yapf: enable

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/apis/inference.py+4-3

	@@ -29,6 +29,7 @@ from mmdet.models import build_detector
29	import cv2	29	import cv2
30	from scipy import ndimage	30	from scipy import ndimage
31		31
		32	+
32	def init_detector(config, checkpoint=None, device='cuda:0'):	33	def init_detector(config, checkpoint=None, device='cuda:0'):
33	"""Initialize a detector from config file.	34	"""Initialize a detector from config file.
34		35
	@@ -280,7 +281,7 @@ def show_result_ins(img,
280	for _ in range(num_mask)	281	for _ in range(num_mask)
281	]	282	]
282	for idx in range(num_mask):	283	for idx in range(num_mask):
283	- idx = -(idx+1)	284	+ idx = -(idx + 1)
284	cur_mask = seg_label[idx, :, :]	285	cur_mask = seg_label[idx, :, :]
285	cur_mask = mmcv.imresize(cur_mask, (w, h))	286	cur_mask = mmcv.imresize(cur_mask, (w, h))
286	cur_mask = (cur_mask > 0.5).astype(np.uint8)	287	cur_mask = (cur_mask > 0.5).astype(np.uint8)
	@@ -293,11 +294,11 @@ def show_result_ins(img,
293	cur_cate = cate_label[idx]	294	cur_cate = cate_label[idx]
294	cur_score = cate_score[idx]	295	cur_score = cate_score[idx]
295	label_text = class_names[cur_cate]	296	label_text = class_names[cur_cate]
296	- #label_text += '\|{:.02f}'.format(cur_score)	297	+ # label_text += '\|{:.02f}'.format(cur_score)
297	center_y, center_x = ndimage.measurements.center_of_mass(cur_mask)	298	center_y, center_x = ndimage.measurements.center_of_mass(cur_mask)
298	vis_pos = (max(int(center_x) - 10, 0), int(center_y))	299	vis_pos = (max(int(center_x) - 10, 0), int(center_y))
299	cv2.putText(img_show, label_text, vis_pos,	300	cv2.putText(img_show, label_text, vis_pos,
300	- cv2.FONT_HERSHEY_COMPLEX, 0.3, (255, 255, 255)) # green	301	+ cv2.FONT_HERSHEY_COMPLEX, 0.3, (255, 255, 255)) # green
301	if out_file is None:	302	if out_file is None:
302	return img_show	303	return img_show
303	else:	304	else:

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/apis/train.py+20-9

	@@ -30,6 +30,8 @@ from mmdet.datasets import DATASETS, build_dataloader
30	from mmdet.utils import get_root_logger	30	from mmdet.utils import get_root_logger
31	from apex import amp	31	from apex import amp
32	import apex	32	import apex
		33	+
		34	+
33	def set_random_seed(seed, deterministic=False):	35	def set_random_seed(seed, deterministic=False):
34	"""Set random seed.	36	"""Set random seed.
35		37
	@@ -103,7 +105,8 @@ def train_detector(model,
103	cfg,	105	cfg,
104	distributed=False,	106	distributed=False,
105	validate=False,	107	validate=False,
106	- timestamp=None):	108	+ timestamp=None,
		109	+ train_performance=False):
107	logger = get_root_logger(cfg.log_level)	110	logger = get_root_logger(cfg.log_level)
108		111
109	# start training	112	# start training
	@@ -114,7 +117,8 @@ def train_detector(model,
114	cfg,	117	cfg,
115	validate=validate,	118	validate=validate,
116	logger=logger,	119	logger=logger,
117	- timestamp=timestamp)	120	+ timestamp=timestamp,
		121	+ train_performance=train_performance)
118	else:	122	else:
119	_non_dist_train(	123	_non_dist_train(
120	model,	124	model,
	@@ -122,7 +126,8 @@ def train_detector(model,
122	cfg,	126	cfg,
123	validate=validate,	127	validate=validate,
124	logger=logger,	128	logger=logger,
125	- timestamp=timestamp)	129	+ timestamp=timestamp,
		130	+ train_performance=train_performance)
126		131
127		132
128	def build_optimizer(model, optimizer_cfg):	133	def build_optimizer(model, optimizer_cfg):
	@@ -210,7 +215,8 @@ def _dist_train(model,
210	cfg,	215	cfg,
211	validate=False,	216	validate=False,
212	logger=None,	217	logger=None,
213	- timestamp=None):	218	+ timestamp=None,
		219	+ train_performance=False):
214	# prepare data loaders	220	# prepare data loaders
215	dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]	221	dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
216	data_loaders = [	222	data_loaders = [
	@@ -237,7 +243,8 @@ def _dist_train(model,
237	# )	243	# )
238	# build runner	244	# build runner
239	runner = Runner(	245	runner = Runner(
240	- model, batch_processor, optimizer, cfg.work_dir, logger=logger, samples_per_gpu=cfg.data.imgs_per_gpu, num_of_gpus = cfg.gpus)	246	+ model, batch_processor, optimizer, cfg.work_dir, logger=logger, samples_per_gpu=cfg.data.imgs_per_gpu,
		247	+ num_of_gpus=cfg.gpus)
241	# an ugly walkaround to make the .log and .log.json filenames the same	248	# an ugly walkaround to make the .log and .log.json filenames the same
242	runner.timestamp = timestamp	249	runner.timestamp = timestamp
243		250
	@@ -284,6 +291,7 @@ def _dist_train(model,
284	runner.resume(cfg.resume_from)	291	runner.resume(cfg.resume_from)
285	elif cfg.load_from:	292	elif cfg.load_from:
286	runner.load_checkpoint(cfg.load_from)	293	runner.load_checkpoint(cfg.load_from)
		294	+ runner.train_performance = train_performance
287	runner.run(data_loaders, cfg.workflow, cfg.total_epochs)	295	runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
288		296
289		297
	@@ -292,7 +300,8 @@ def _non_dist_train(model,
292	cfg,	300	cfg,
293	validate=False,	301	validate=False,
294	logger=None,	302	logger=None,
295	- timestamp=None):	303	+ timestamp=None,
		304	+ train_performance=False):
296	if validate:	305	if validate:
297	raise NotImplementedError('Built-in validation is not implemented '	306	raise NotImplementedError('Built-in validation is not implemented '
298	'yet in not-distributed training. Use '	307	'yet in not-distributed training. Use '
	@@ -318,9 +327,10 @@ def _non_dist_train(model,
318	model = MMDataParallel(model.npu(), device_ids=range(cfg.gpus))	327	model = MMDataParallel(model.npu(), device_ids=range(cfg.gpus))
319		328
320	# build runner	329	# build runner
321	- #optimizer = build_optimizer(model, cfg.optimizer)	330	+ # optimizer = build_optimizer(model, cfg.optimizer)
322	runner = Runner(	331	runner = Runner(
323	- model, batch_processor, optimizer, cfg.work_dir, logger=logger, samples_per_gpu=cfg.data.imgs_per_gpu, num_of_gpus = cfg.gpus)	332	+ model, batch_processor, optimizer, cfg.work_dir, logger=logger, samples_per_gpu=cfg.data.imgs_per_gpu,
		333	+ num_of_gpus=cfg.gpus)
324	# an ugly walkaround to make the .log and .log.json filenames the same	334	# an ugly walkaround to make the .log and .log.json filenames the same
325	runner.timestamp = timestamp	335	runner.timestamp = timestamp
326	# fp16 setting	336	# fp16 setting
	@@ -337,4 +347,5 @@ def _non_dist_train(model,
337	runner.resume(cfg.resume_from)	347	runner.resume(cfg.resume_from)
338	elif cfg.load_from:	348	elif cfg.load_from:
339	runner.load_checkpoint(cfg.load_from)	349	runner.load_checkpoint(cfg.load_from)
340	- runner.run(data_loaders, cfg.workflow, cfg.total_epochs)	350	+ runner.train_performance = train_performance
		351	+ runner.run(data_loaders, cfg.workflow, cfg.total_epochs)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/anchor/anchor_generator.py+2-2

	@@ -107,6 +107,6 @@ class AnchorGenerator(object):
107	valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)	107	valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)
108	valid = valid_xx & valid_yy	108	valid = valid_xx & valid_yy
109	valid = valid[:,	109	valid = valid[:,
110	- None].expand(valid.size(0),	110	+ None].expand(valid.size(0),
111	- self.num_base_anchors).contiguous().view(-1)	111	+ self.num_base_anchors).contiguous().view(-1)
112	return valid	112	return valid

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/anchor/anchor_target.py+20-20

	@@ -62,19 +62,19 @@ def anchor_target(anchor_list,
62	gt_labels_list = [None for _ in range(num_imgs)]	62	gt_labels_list = [None for _ in range(num_imgs)]
63	(all_labels, all_label_weights, all_bbox_targets, all_bbox_weights,	63	(all_labels, all_label_weights, all_bbox_targets, all_bbox_weights,
64	pos_inds_list, neg_inds_list) = multi_apply(	64	pos_inds_list, neg_inds_list) = multi_apply(
65	- anchor_target_single,	65	+ anchor_target_single,
66	- anchor_list,	66	+ anchor_list,
67	- valid_flag_list,	67	+ valid_flag_list,
68	- gt_bboxes_list,	68	+ gt_bboxes_list,
69	- gt_bboxes_ignore_list,	69	+ gt_bboxes_ignore_list,
70	- gt_labels_list,	70	+ gt_labels_list,
71	- img_metas,	71	+ img_metas,
72	- target_means=target_means,	72	+ target_means=target_means,
73	- target_stds=target_stds,	73	+ target_stds=target_stds,
74	- cfg=cfg,	74	+ cfg=cfg,
75	- label_channels=label_channels,	75	+ label_channels=label_channels,
76	- sampling=sampling,	76	+ sampling=sampling,
77	- unmap_outputs=unmap_outputs)	77	+ unmap_outputs=unmap_outputs)
78	# no valid anchors	78	# no valid anchors
79	if any([labels is None for labels in all_labels]):	79	if any([labels is None for labels in all_labels]):
80	return None	80	return None
	@@ -121,7 +121,7 @@ def anchor_target_single(flat_anchors,
121	img_meta['img_shape'][:2],	121	img_meta['img_shape'][:2],
122	cfg.allowed_border)	122	cfg.allowed_border)
123	if not inside_flags.any():	123	if not inside_flags.any():
124	- return (None, ) * 6	124	+ return (None,) * 6
125	# assign gt and sample anchors	125	# assign gt and sample anchors
126	anchors = flat_anchors[inside_flags, :]	126	anchors = flat_anchors[inside_flags, :]
127		127
	@@ -180,10 +180,10 @@ def anchor_inside_flags(flat_anchors,
180	img_h, img_w = img_shape[:2]	180	img_h, img_w = img_shape[:2]
181	if allowed_border >= 0:	181	if allowed_border >= 0:
182	inside_flags = valid_flags & \	182	inside_flags = valid_flags & \
183	- (flat_anchors[:, 0] >= -allowed_border).type(torch.uint8) & \	183	+ (flat_anchors[:, 0] >= -allowed_border).type(torch.uint8) & \
184	- (flat_anchors[:, 1] >= -allowed_border).type(torch.uint8) & \	184	+ (flat_anchors[:, 1] >= -allowed_border).type(torch.uint8) & \
185	- (flat_anchors[:, 2] < img_w + allowed_border).type(torch.uint8) & \	185	+ (flat_anchors[:, 2] < img_w + allowed_border).type(torch.uint8) & \
186	- (flat_anchors[:, 3] < img_h + allowed_border).type(torch.uint8)	186	+ (flat_anchors[:, 3] < img_h + allowed_border).type(torch.uint8)
187	else:	187	else:
188	inside_flags = valid_flags	188	inside_flags = valid_flags
189	return inside_flags	189	return inside_flags
	@@ -193,10 +193,10 @@ def unmap(data, count, inds, fill=0):
193	""" Unmap a subset of item (data) back to the original set of items (of	193	""" Unmap a subset of item (data) back to the original set of items (of
194	size count) """	194	size count) """
195	if data.dim() == 1:	195	if data.dim() == 1:
196	- ret = data.new_full((count, ), fill)	196	+ ret = data.new_full((count,), fill)
197	ret[inds] = data	197	ret[inds] = data
198	else:	198	else:
199	- new_size = (count, ) + data.size()[1:]	199	+ new_size = (count,) + data.size()[1:]
200	ret = data.new_full(new_size, fill)	200	ret = data.new_full(new_size, fill)
201	ret[inds, :] = data	201	ret[inds, :] = data
202	return ret	202	return ret

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/anchor/guided_anchor_target.py+18-18

	@@ -93,7 +93,7 @@ def ga_loc_target(gt_bboxes_list,
93	scale = torch.sqrt((gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) *	93	scale = torch.sqrt((gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) *
94	(gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1))	94	(gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1))
95	min_anchor_size = scale.new_full(	95	min_anchor_size = scale.new_full(
96	- (1, ), float(anchor_scale * anchor_strides[0]))	96	+ (1,), float(anchor_scale * anchor_strides[0]))
97	# assign gt bboxes to different feature levels w.r.t. their scales	97	# assign gt bboxes to different feature levels w.r.t. their scales
98	target_lvls = torch.floor(	98	target_lvls = torch.floor(
99	torch.log2(scale) - torch.log2(min_anchor_size) + 0.5)	99	torch.log2(scale) - torch.log2(min_anchor_size) + 0.5)
	@@ -109,11 +109,11 @@ def ga_loc_target(gt_bboxes_list,
109	ctr_x1, ctr_y1, ctr_x2, ctr_y2 = calc_region(	109	ctr_x1, ctr_y1, ctr_x2, ctr_y2 = calc_region(
110	gt_, r1, featmap_sizes[lvl])	110	gt_, r1, featmap_sizes[lvl])
111	all_loc_targets[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,	111	all_loc_targets[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
112	- ctr_x1:ctr_x2 + 1] = 1	112	+ ctr_x1:ctr_x2 + 1] = 1
113	all_loc_weights[lvl][img_id, 0, ignore_y1:ignore_y2 + 1,	113	all_loc_weights[lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
114	- ignore_x1:ignore_x2 + 1] = 0	114	+ ignore_x1:ignore_x2 + 1] = 0
115	all_loc_weights[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,	115	all_loc_weights[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
116	- ctr_x1:ctr_x2 + 1] = 1	116	+ ctr_x1:ctr_x2 + 1] = 1
117	# calculate ignore map on nearby low level feature	117	# calculate ignore map on nearby low level feature
118	if lvl > 0:	118	if lvl > 0:
119	d_lvl = lvl - 1	119	d_lvl = lvl - 1
	@@ -122,7 +122,7 @@ def ga_loc_target(gt_bboxes_list,
122	ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(	122	ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
123	gt_, r2, featmap_sizes[d_lvl])	123	gt_, r2, featmap_sizes[d_lvl])
124	all_ignore_map[d_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,	124	all_ignore_map[d_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
125	- ignore_x1:ignore_x2 + 1] = 1	125	+ ignore_x1:ignore_x2 + 1] = 1
126	# calculate ignore map on nearby high level feature	126	# calculate ignore map on nearby high level feature
127	if lvl < num_lvls - 1:	127	if lvl < num_lvls - 1:
128	u_lvl = lvl + 1	128	u_lvl = lvl + 1
	@@ -131,7 +131,7 @@ def ga_loc_target(gt_bboxes_list,
131	ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(	131	ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
132	gt_, r2, featmap_sizes[u_lvl])	132	gt_, r2, featmap_sizes[u_lvl])
133	all_ignore_map[u_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,	133	all_ignore_map[u_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
134	- ignore_x1:ignore_x2 + 1] = 1	134	+ ignore_x1:ignore_x2 + 1] = 1
135	for lvl_id in range(num_lvls):	135	for lvl_id in range(num_lvls):
136	# ignore negative regions w.r.t. ignore map	136	# ignore negative regions w.r.t. ignore map
137	all_loc_weights[lvl_id][(all_loc_weights[lvl_id] < 0)	137	all_loc_weights[lvl_id][(all_loc_weights[lvl_id] < 0)
	@@ -191,17 +191,17 @@ def ga_shape_target(approx_list,
191	gt_bboxes_ignore_list = [None for _ in range(num_imgs)]	191	gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
192	(all_bbox_anchors, all_bbox_gts, all_bbox_weights, pos_inds_list,	192	(all_bbox_anchors, all_bbox_gts, all_bbox_weights, pos_inds_list,
193	neg_inds_list) = multi_apply(	193	neg_inds_list) = multi_apply(
194	- ga_shape_target_single,	194	+ ga_shape_target_single,
195	- approx_flat_list,	195	+ approx_flat_list,
196	- inside_flag_flat_list,	196	+ inside_flag_flat_list,
197	- square_flat_list,	197	+ square_flat_list,
198	- gt_bboxes_list,	198	+ gt_bboxes_list,
199	- gt_bboxes_ignore_list,	199	+ gt_bboxes_ignore_list,
200	- img_metas,	200	+ img_metas,
201	- approxs_per_octave=approxs_per_octave,	201	+ approxs_per_octave=approxs_per_octave,
202	- cfg=cfg,	202	+ cfg=cfg,
203	- sampling=sampling,	203	+ sampling=sampling,
204	- unmap_outputs=unmap_outputs)	204	+ unmap_outputs=unmap_outputs)
205	# no valid anchors	205	# no valid anchors
206	if any([bbox_anchors is None for bbox_anchors in all_bbox_anchors]):	206	if any([bbox_anchors is None for bbox_anchors in all_bbox_anchors]):
207	return None	207	return None
	@@ -264,7 +264,7 @@ def ga_shape_target_single(flat_approxs,
264	tuple	264	tuple
265	"""	265	"""
266	if not inside_flags.any():	266	if not inside_flags.any():
267	- return (None, ) * 5	267	+ return (None,) * 5
268	# assign gt and sample anchors	268	# assign gt and sample anchors
269	expand_inside_flags = inside_flags[:, None].expand(	269	expand_inside_flags = inside_flags[:, None].expand(
270	-1, approxs_per_octave).reshape(-1)	270	-1, approxs_per_octave).reshape(-1)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/anchor/point_generator.py+1-1

	@@ -30,7 +30,7 @@ class PointGenerator(object):
30	shift_x = torch.arange(0., feat_w, device=device) * stride	30	shift_x = torch.arange(0., feat_w, device=device) * stride
31	shift_y = torch.arange(0., feat_h, device=device) * stride	31	shift_y = torch.arange(0., feat_h, device=device) * stride
32	shift_xx, shift_yy = self._meshgrid(shift_x, shift_y)	32	shift_xx, shift_yy = self._meshgrid(shift_x, shift_y)
33	- stride = shift_x.new_full((shift_xx.shape[0], ), stride)	33	+ stride = shift_x.new_full((shift_xx.shape[0],), stride)
34	shifts = torch.stack([shift_xx, shift_yy, stride], dim=-1)	34	shifts = torch.stack([shift_xx, shift_yy, stride], dim=-1)
35	all_points = shifts.to(device)	35	all_points = shifts.to(device)
36	return all_points	36	return all_points

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/anchor/point_target.py+13-13

	@@ -59,16 +59,16 @@ def point_target(proposals_list,
59	gt_labels_list = [None for _ in range(num_imgs)]	59	gt_labels_list = [None for _ in range(num_imgs)]
60	(all_labels, all_label_weights, all_bbox_gt, all_proposals,	60	(all_labels, all_label_weights, all_bbox_gt, all_proposals,
61	all_proposal_weights, pos_inds_list, neg_inds_list) = multi_apply(	61	all_proposal_weights, pos_inds_list, neg_inds_list) = multi_apply(
62	- point_target_single,	62	+ point_target_single,
63	- proposals_list,	63	+ proposals_list,
64	- valid_flag_list,	64	+ valid_flag_list,
65	- gt_bboxes_list,	65	+ gt_bboxes_list,
66	- gt_bboxes_ignore_list,	66	+ gt_bboxes_ignore_list,
67	- gt_labels_list,	67	+ gt_labels_list,
68	- cfg=cfg,	68	+ cfg=cfg,
69	- label_channels=label_channels,	69	+ label_channels=label_channels,
70	- sampling=sampling,	70	+ sampling=sampling,
71	- unmap_outputs=unmap_outputs)	71	+ unmap_outputs=unmap_outputs)
72	# no valid points	72	# no valid points
73	if any([labels is None for labels in all_labels]):	73	if any([labels is None for labels in all_labels]):
74	return None	74	return None
	@@ -112,7 +112,7 @@ def point_target_single(flat_proposals,
112	unmap_outputs=True):	112	unmap_outputs=True):
113	inside_flags = valid_flags	113	inside_flags = valid_flags
114	if not inside_flags.any():	114	if not inside_flags.any():
115	- return (None, ) * 7	115	+ return (None,) * 7
116	# assign gt and sample proposals	116	# assign gt and sample proposals
117	proposals = flat_proposals[inside_flags, :]	117	proposals = flat_proposals[inside_flags, :]
118		118
	@@ -170,10 +170,10 @@ def unmap(data, count, inds, fill=0):
170	""" Unmap a subset of item (data) back to the original set of items (of	170	""" Unmap a subset of item (data) back to the original set of items (of
171	size count) """	171	size count) """
172	if data.dim() == 1:	172	if data.dim() == 1:
173	- ret = data.new_full((count, ), fill)	173	+ ret = data.new_full((count,), fill)
174	ret[inds] = data	174	ret[inds] = data
175	else:	175	else:
176	- new_size = (count, ) + data.size()[1:]	176	+ new_size = (count,) + data.size()[1:]
177	ret = data.new_full(new_size, fill)	177	ret = data.new_full(new_size, fill)
178	ret[inds, :] = data	178	ret[inds, :] = data
179	return ret	179	return ret

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/assigners/approx_max_iou_assigner.py+1-1

	@@ -114,7 +114,7 @@ class ApproxMaxIoUAssigner(MaxIoUAssigner):
114	approxs.view(num_squares, approxs_per_octave, 4), 0,	114	approxs.view(num_squares, approxs_per_octave, 4), 0,
115	1).contiguous().view(-1, 4)	115	1).contiguous().view(-1, 4)
116	assign_on_cpu = True if (self.gpu_assign_thr > 0) and (	116	assign_on_cpu = True if (self.gpu_assign_thr > 0) and (
117	- num_gts > self.gpu_assign_thr) else False	117	+ num_gts > self.gpu_assign_thr) else False
118	# compute overlap and assign gt on CPU when number of GT is large	118	# compute overlap and assign gt on CPU when number of GT is large
119	if assign_on_cpu:	119	if assign_on_cpu:
120	device = approxs.device	120	device = approxs.device

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/assigners/atss_assigner.py+4-4

	@@ -78,20 +78,20 @@ class ATSSAssigner(BaseAssigner):
78	overlaps = bbox_overlaps(bboxes, gt_bboxes)	78	overlaps = bbox_overlaps(bboxes, gt_bboxes)
79		79
80	# assign 0 by default	80	# assign 0 by default
81	- assigned_gt_inds = overlaps.new_full((num_bboxes, ),	81	+ assigned_gt_inds = overlaps.new_full((num_bboxes,),
82	0,	82	0,
83	dtype=torch.long)	83	dtype=torch.long)
84		84
85	if num_gt == 0 or num_bboxes == 0:	85	if num_gt == 0 or num_bboxes == 0:
86	# No ground truth or boxes, return empty assignment	86	# No ground truth or boxes, return empty assignment
87	- max_overlaps = overlaps.new_zeros((num_bboxes, ))	87	+ max_overlaps = overlaps.new_zeros((num_bboxes,))
88	if num_gt == 0:	88	if num_gt == 0:
89	# No truth, assign everything to background	89	# No truth, assign everything to background
90	assigned_gt_inds[:] = 0	90	assigned_gt_inds[:] = 0
91	if gt_labels is None:	91	if gt_labels is None:
92	assigned_labels = None	92	assigned_labels = None
93	else:	93	else:
94	- assigned_labels = overlaps.new_zeros((num_bboxes, ),	94	+ assigned_labels = overlaps.new_zeros((num_bboxes,),
95	dtype=torch.long)	95	dtype=torch.long)
96	return AssignResult(	96	return AssignResult(
97	num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)	97	num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)
	@@ -162,7 +162,7 @@ class ATSSAssigner(BaseAssigner):
162	max_overlaps != -INF] = argmax_overlaps[max_overlaps != -INF] + 1	162	max_overlaps != -INF] = argmax_overlaps[max_overlaps != -INF] + 1
163		163
164	if gt_labels is not None:	164	if gt_labels is not None:
165	- assigned_labels = assigned_gt_inds.new_zeros((num_bboxes, ))	165	+ assigned_labels = assigned_gt_inds.new_zeros((num_bboxes,))
166	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()	166	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()
167	if pos_inds.numel() > 0:	167	if pos_inds.numel() > 0:
168	assigned_labels[pos_inds] = gt_labels[	168	assigned_labels[pos_inds] = gt_labels[

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/assigners/max_iou_assigner.py+5-5

	@@ -98,7 +98,7 @@ class MaxIoUAssigner(BaseAssigner):
98	>>> assert torch.all(assign_result.gt_inds == expected_gt_inds)	98	>>> assert torch.all(assign_result.gt_inds == expected_gt_inds)
99	"""	99	"""
100	assign_on_cpu = True if (self.gpu_assign_thr > 0) and (	100	assign_on_cpu = True if (self.gpu_assign_thr > 0) and (
101	- gt_bboxes.shape[0] > self.gpu_assign_thr) else False	101	+ gt_bboxes.shape[0] > self.gpu_assign_thr) else False
102	# compute overlap and assign gt on CPU when number of GT is large	102	# compute overlap and assign gt on CPU when number of GT is large
103	if assign_on_cpu:	103	if assign_on_cpu:
104	device = bboxes.device	104	device = bboxes.device
	@@ -146,20 +146,20 @@ class MaxIoUAssigner(BaseAssigner):
146	num_gts, num_bboxes = overlaps.size(0), overlaps.size(1)	146	num_gts, num_bboxes = overlaps.size(0), overlaps.size(1)
147		147
148	# 1. assign -1 by default	148	# 1. assign -1 by default
149	- assigned_gt_inds = overlaps.new_full((num_bboxes, ),	149	+ assigned_gt_inds = overlaps.new_full((num_bboxes,),
150	-1,	150	-1,
151	dtype=torch.long)	151	dtype=torch.long)
152		152
153	if num_gts == 0 or num_bboxes == 0:	153	if num_gts == 0 or num_bboxes == 0:
154	# No ground truth or boxes, return empty assignment	154	# No ground truth or boxes, return empty assignment
155	- max_overlaps = overlaps.new_zeros((num_bboxes, ))	155	+ max_overlaps = overlaps.new_zeros((num_bboxes,))
156	if num_gts == 0:	156	if num_gts == 0:
157	# No truth, assign everything to background	157	# No truth, assign everything to background
158	assigned_gt_inds[:] = 0	158	assigned_gt_inds[:] = 0
159	if gt_labels is None:	159	if gt_labels is None:
160	assigned_labels = None	160	assigned_labels = None
161	else:	161	else:
162	- assigned_labels = overlaps.new_zeros((num_bboxes, ),	162	+ assigned_labels = overlaps.new_zeros((num_bboxes,),
163	dtype=torch.long)	163	dtype=torch.long)
164	return AssignResult(	164	return AssignResult(
165	num_gts,	165	num_gts,
	@@ -197,7 +197,7 @@ class MaxIoUAssigner(BaseAssigner):
197	assigned_gt_inds[gt_argmax_overlaps[i]] = i + 1	197	assigned_gt_inds[gt_argmax_overlaps[i]] = i + 1
198		198
199	if gt_labels is not None:	199	if gt_labels is not None:
200	- assigned_labels = assigned_gt_inds.new_zeros((num_bboxes, ))	200	+ assigned_labels = assigned_gt_inds.new_zeros((num_bboxes,))
201	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()	201	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()
202	if pos_inds.numel() > 0:	202	if pos_inds.numel() > 0:
203	assigned_labels[pos_inds] = gt_labels[	203	assigned_labels[pos_inds] = gt_labels[

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/assigners/point_assigner.py+5-5

	@@ -65,13 +65,13 @@ class PointAssigner(BaseAssigner):
65		65
66	if num_gts == 0 or num_points == 0:	66	if num_gts == 0 or num_points == 0:
67	# If no truth assign everything to the background	67	# If no truth assign everything to the background
68	- assigned_gt_inds = points.new_full((num_points, ),	68	+ assigned_gt_inds = points.new_full((num_points,),
69	0,	69	0,
70	dtype=torch.long)	70	dtype=torch.long)
71	if gt_labels is None:	71	if gt_labels is None:
72	assigned_labels = None	72	assigned_labels = None
73	else:	73	else:
74	- assigned_labels = points.new_zeros((num_points, ),	74	+ assigned_labels = points.new_zeros((num_points,),
75	dtype=torch.long)	75	dtype=torch.long)
76	return AssignResult(	76	return AssignResult(
77	num_gts, assigned_gt_inds, None, labels=assigned_labels)	77	num_gts, assigned_gt_inds, None, labels=assigned_labels)
	@@ -91,9 +91,9 @@ class PointAssigner(BaseAssigner):
91	gt_bboxes_lvl = torch.clamp(gt_bboxes_lvl, min=lvl_min, max=lvl_max)	91	gt_bboxes_lvl = torch.clamp(gt_bboxes_lvl, min=lvl_min, max=lvl_max)
92		92
93	# stores the assigned gt index of each point	93	# stores the assigned gt index of each point
94	- assigned_gt_inds = points.new_zeros((num_points, ), dtype=torch.long)	94	+ assigned_gt_inds = points.new_zeros((num_points,), dtype=torch.long)
95	# stores the assigned gt dist (to this point) of each point	95	# stores the assigned gt dist (to this point) of each point
96	- assigned_gt_dist = points.new_full((num_points, ), float('inf'))	96	+ assigned_gt_dist = points.new_full((num_points,), float('inf'))
97	points_range = torch.arange(points.shape[0])	97	points_range = torch.arange(points.shape[0])
98		98
99	for idx in range(num_gts):	99	for idx in range(num_gts):
	@@ -132,7 +132,7 @@ class PointAssigner(BaseAssigner):
132	less_than_recorded_index]	132	less_than_recorded_index]
133		133
134	if gt_labels is not None:	134	if gt_labels is not None:
135	- assigned_labels = assigned_gt_inds.new_zeros((num_points, ))	135	+ assigned_labels = assigned_gt_inds.new_zeros((num_points,))
136	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()	136	pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()
137	if pos_inds.numel() > 0:	137	if pos_inds.numel() > 0:
138	assigned_labels[pos_inds] = gt_labels[	138	assigned_labels[pos_inds] = gt_labels[

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/geometry.py+4-4

	@@ -75,11 +75,11 @@ def bbox_overlaps(bboxes1, bboxes2, mode='iou', is_aligned=False):
75	wh = (rb - lt + 1).clamp(min=0) # [rows, 2]	75	wh = (rb - lt + 1).clamp(min=0) # [rows, 2]
76	overlap = wh[:, 0] * wh[:, 1]	76	overlap = wh[:, 0] * wh[:, 1]
77	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (	77	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (
78	- bboxes1[:, 3] - bboxes1[:, 1] + 1)	78	+ bboxes1[:, 3] - bboxes1[:, 1] + 1)
79		79
80	if mode == 'iou':	80	if mode == 'iou':
81	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (	81	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (
82	- bboxes2[:, 3] - bboxes2[:, 1] + 1)	82	+ bboxes2[:, 3] - bboxes2[:, 1] + 1)
83	ious = overlap / (area1 + area2 - overlap)	83	ious = overlap / (area1 + area2 - overlap)
84	else:	84	else:
85	ious = overlap / area1	85	ious = overlap / area1
	@@ -90,11 +90,11 @@ def bbox_overlaps(bboxes1, bboxes2, mode='iou', is_aligned=False):
90	wh = (rb - lt + 1).clamp(min=0) # [rows, cols, 2]	90	wh = (rb - lt + 1).clamp(min=0) # [rows, cols, 2]
91	overlap = wh[:, :, 0] * wh[:, :, 1]	91	overlap = wh[:, :, 0] * wh[:, :, 1]
92	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (	92	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (
93	- bboxes1[:, 3] - bboxes1[:, 1] + 1)	93	+ bboxes1[:, 3] - bboxes1[:, 1] + 1)
94		94
95	if mode == 'iou':	95	if mode == 'iou':
96	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (	96	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (
97	- bboxes2[:, 3] - bboxes2[:, 1] + 1)	97	+ bboxes2[:, 3] - bboxes2[:, 1] + 1)
98	ious = overlap / (area1[:, None] + area2 - overlap)	98	ious = overlap / (area1[:, None] + area2 - overlap)
99	else:	99	else:
100	ious = overlap / (area1[:, None])	100	ious = overlap / (area1[:, None])

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/bbox/samplers/base_sampler.py+1-1

	@@ -80,7 +80,7 @@ class BaseSampler(metaclass=ABCMeta):
80		80
81	bboxes = bboxes[:, :4]	81	bboxes = bboxes[:, :4]
82		82
83	- gt_flags = bboxes.new_zeros((bboxes.shape[0], ), dtype=torch.uint8)	83	+ gt_flags = bboxes.new_zeros((bboxes.shape[0],), dtype=torch.uint8)
84	if self.add_gt_as_proposals and len(gt_bboxes) > 0:	84	if self.add_gt_as_proposals and len(gt_bboxes) > 0:
85	if gt_labels is None:	85	if gt_labels is None:
86	raise ValueError(	86	raise ValueError(

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/evaluation/bbox_overlaps.py+2-2

	@@ -43,9 +43,9 @@ def bbox_overlaps(bboxes1, bboxes2, mode='iou'):
43	ious = np.zeros((cols, rows), dtype=np.float32)	43	ious = np.zeros((cols, rows), dtype=np.float32)
44	exchange = True	44	exchange = True
45	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (	45	area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (
46	- bboxes1[:, 3] - bboxes1[:, 1] + 1)	46	+ bboxes1[:, 3] - bboxes1[:, 1] + 1)
47	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (	47	area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (
48	- bboxes2[:, 3] - bboxes2[:, 1] + 1)	48	+ bboxes2[:, 3] - bboxes2[:, 1] + 1)
49	for i in range(bboxes1.shape[0]):	49	for i in range(bboxes1.shape[0]):
50	x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0])	50	x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0])
51	y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1])	51	y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1])

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/evaluation/coco_utils.py+1-1

	@@ -105,7 +105,7 @@ def fast_eval_recall(results,
105	elif not isinstance(results, list):	105	elif not isinstance(results, list):
106	raise TypeError(	106	raise TypeError(
107	'results must be a list of numpy arrays or a filename, not {}'.	107	'results must be a list of numpy arrays or a filename, not {}'.
108	- format(type(results)))	108	+ format(type(results)))
109		109
110	gt_bboxes = []	110	gt_bboxes = []
111	img_ids = coco.getImgIds()	111	img_ids = coco.getImgIds()

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/evaluation/mean_ap.py+5-5

	@@ -113,7 +113,7 @@ def tpfp_imagenet(det_bboxes,
113	fp[...] = 1	113	fp[...] = 1
114	else:	114	else:
115	det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0] + 1) * (	115	det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0] + 1) * (
116	- det_bboxes[:, 3] - det_bboxes[:, 1] + 1)	116	+ det_bboxes[:, 3] - det_bboxes[:, 1] + 1)
117	for i, (min_area, max_area) in enumerate(area_ranges):	117	for i, (min_area, max_area) in enumerate(area_ranges):
118	fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1	118	fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
119	return tp, fp	119	return tp, fp
	@@ -209,7 +209,7 @@ def tpfp_default(det_bboxes,
209	fp[...] = 1	209	fp[...] = 1
210	else:	210	else:
211	det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0] + 1) * (	211	det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0] + 1) * (
212	- det_bboxes[:, 3] - det_bboxes[:, 1] + 1)	212	+ det_bboxes[:, 3] - det_bboxes[:, 1] + 1)
213	for i, (min_area, max_area) in enumerate(area_ranges):	213	for i, (min_area, max_area) in enumerate(area_ranges):
214	fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1	214	fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
215	return tp, fp	215	return tp, fp
	@@ -228,7 +228,7 @@ def tpfp_default(det_bboxes,
228	gt_area_ignore = np.zeros_like(gt_ignore_inds, dtype=bool)	228	gt_area_ignore = np.zeros_like(gt_ignore_inds, dtype=bool)
229	else:	229	else:
230	gt_areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) * (	230	gt_areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) * (
231	- gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1)	231	+ gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1)
232	gt_area_ignore = (gt_areas < min_area) \| (gt_areas >= max_area)	232	gt_area_ignore = (gt_areas < min_area) \| (gt_areas >= max_area)
233	for i in sort_inds:	233	for i in sort_inds:
234	if ious_max[i] >= iou_thr:	234	if ious_max[i] >= iou_thr:
	@@ -318,7 +318,7 @@ def eval_map(det_results,
318	num_imgs = len(det_results)	318	num_imgs = len(det_results)
319	num_scales = len(scale_ranges) if scale_ranges is not None else 1	319	num_scales = len(scale_ranges) if scale_ranges is not None else 1
320	num_classes = len(det_results[0]) # positive class num	320	num_classes = len(det_results[0]) # positive class num
321	- area_ranges = ([(rg[0]2, rg[1]2) for rg in scale_ranges]	321	+ area_ranges = ([(rg[0] 2, rg[1] 2) for rg in scale_ranges]
322	if scale_ranges is not None else None)	322	if scale_ranges is not None else None)
323		323
324	pool = Pool(nproc)	324	pool = Pool(nproc)
	@@ -347,7 +347,7 @@ def eval_map(det_results,
347	num_gts[0] += bbox.shape[0]	347	num_gts[0] += bbox.shape[0]
348	else:	348	else:
349	gt_areas = (bbox[:, 2] - bbox[:, 0] + 1) * (	349	gt_areas = (bbox[:, 2] - bbox[:, 0] + 1) * (
350	- bbox[:, 3] - bbox[:, 1] + 1)	350	+ bbox[:, 3] - bbox[:, 1] + 1)
351	for k, (min_area, max_area) in enumerate(area_ranges):	351	for k, (min_area, max_area) in enumerate(area_ranges):
352	num_gts[k] += np.sum((gt_areas >= min_area)	352	num_gts[k] += np.sum((gt_areas >= min_area)
353	& (gt_areas < max_area))	353	& (gt_areas < max_area))

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/evaluation/recall.py+0-1

	@@ -19,7 +19,6 @@ from .bbox_overlaps import bbox_overlaps
19		19
20		20
21	def _recalls(all_ious, proposal_nums, thrs):	21	def _recalls(all_ious, proposal_nums, thrs):
22	-
23	img_num = all_ious.shape[0]	22	img_num = all_ious.shape[0]
24	total_gt_num = sum([ious.shape[0] for ious in all_ious])	23	total_gt_num = sum([ious.shape[0] for ious in all_ious])
25		24

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/mask/mask_target.py+1-1

	@@ -51,5 +51,5 @@ def mask_target_single(pos_proposals, pos_assigned_gt_inds, gt_masks, cfg):
51	mask_targets = torch.from_numpy(np.stack(mask_targets)).float().to(	51	mask_targets = torch.from_numpy(np.stack(mask_targets)).float().to(
52	pos_proposals.device)	52	pos_proposals.device)
53	else:	53	else:
54	- mask_targets = pos_proposals.new_zeros((0, ) + mask_size)	54	+ mask_targets = pos_proposals.new_zeros((0,) + mask_size)
55	return mask_targets	55	return mask_targets

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/post_processing/bbox_nms.py+3-2

	@@ -14,6 +14,7 @@
14		14
15	import torch	15	import torch
16		16
		17	+
17	# from mmdet.ops.nms import nms_wrapper	18	# from mmdet.ops.nms import nms_wrapper
18		19
19		20
	@@ -60,7 +61,7 @@ def multiclass_nms(multi_bboxes,
60	_scores *= score_factors[cls_inds]	61	_scores *= score_factors[cls_inds]
61	cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1)	62	cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1)
62	cls_dets, _ = nms_op(cls_dets, **nms_cfg_)	63	cls_dets, _ = nms_op(cls_dets, **nms_cfg_)
63	- cls_labels = multi_bboxes.new_full((cls_dets.shape[0], ),	64	+ cls_labels = multi_bboxes.new_full((cls_dets.shape[0],),
64	i - 1,	65	i - 1,
65	dtype=torch.long)	66	dtype=torch.long)
66	bboxes.append(cls_dets)	67	bboxes.append(cls_dets)
	@@ -75,6 +76,6 @@ def multiclass_nms(multi_bboxes,
75	labels = labels[inds]	76	labels = labels[inds]
76	else:	77	else:
77	bboxes = multi_bboxes.new_zeros((0, 5))	78	bboxes = multi_bboxes.new_zeros((0, 5))
78	- labels = multi_bboxes.new_zeros((0, ), dtype=torch.long)	79	+ labels = multi_bboxes.new_zeros((0,), dtype=torch.long)
79		80
80	return bboxes, labels	81	return bboxes, labels

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/post_processing/matrix_nms.py+3-3

	@@ -58,7 +58,7 @@ def matrix_nms(seg_masks, cate_labels, cate_scores, kernel='gaussian', sigma=2.0
58	compensate_matrix = torch.exp(-1 * sigma * (compensate_iou ** 2))	58	compensate_matrix = torch.exp(-1 * sigma * (compensate_iou ** 2))
59	decay_coefficient, _ = (decay_matrix / compensate_matrix).min(0)	59	decay_coefficient, _ = (decay_matrix / compensate_matrix).min(0)
60	elif kernel == 'linear':	60	elif kernel == 'linear':
61	- decay_matrix = (1-decay_iou)/(1-compensate_iou)	61	+ decay_matrix = (1 - decay_iou) / (1 - compensate_iou)
62	decay_coefficient, _ = decay_matrix.min(0)	62	decay_coefficient, _ = decay_matrix.min(0)
63	else:	63	else:
64	raise NotImplementedError	64	raise NotImplementedError
	@@ -111,7 +111,7 @@ def multiclass_nms(multi_bboxes,
111	_scores *= score_factors[cls_inds]	111	_scores *= score_factors[cls_inds]
112	cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1)	112	cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1)
113	cls_dets, _ = nms_op(cls_dets, **nms_cfg_)	113	cls_dets, _ = nms_op(cls_dets, **nms_cfg_)
114	- cls_labels = multi_bboxes.new_full((cls_dets.shape[0], ),	114	+ cls_labels = multi_bboxes.new_full((cls_dets.shape[0],),
115	i - 1,	115	i - 1,
116	dtype=torch.long)	116	dtype=torch.long)
117	bboxes.append(cls_dets)	117	bboxes.append(cls_dets)
	@@ -126,6 +126,6 @@ def multiclass_nms(multi_bboxes,
126	labels = labels[inds]	126	labels = labels[inds]
127	else:	127	else:
128	bboxes = multi_bboxes.new_zeros((0, 5))	128	bboxes = multi_bboxes.new_zeros((0, 5))
129	- labels = multi_bboxes.new_zeros((0, ), dtype=torch.long)	129	+ labels = multi_bboxes.new_zeros((0,), dtype=torch.long)
130		130
131	return bboxes, labels	131	return bboxes, labels

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/core/utils/misc.py+2-2

	@@ -42,10 +42,10 @@ def unmap(data, count, inds, fill=0):
42	""" Unmap a subset of item (data) back to the original set of items (of	42	""" Unmap a subset of item (data) back to the original set of items (of
43	size count) """	43	size count) """
44	if data.dim() == 1:	44	if data.dim() == 1:
45	- ret = data.new_full((count, ), fill)	45	+ ret = data.new_full((count,), fill)
46	ret[inds] = data	46	ret[inds] = data
47	else:	47	else:
48	- new_size = (count, ) + data.size()[1:]	48	+ new_size = (count,) + data.size()[1:]
49	ret = data.new_full(new_size, fill)	49	ret = data.new_full(new_size, fill)
50	ret[inds, :] = data	50	ret[inds, :] = data
51	return ret	51	return ret

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/cityscapes.py+0-1

	@@ -18,6 +18,5 @@ from .registry import DATASETS
18		18
19	@DATASETS.register_module	19	@DATASETS.register_module
20	class CityscapesDataset(CocoDataset):	20	class CityscapesDataset(CocoDataset):
21	-
22	CLASSES = ('person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',	21	CLASSES = ('person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',
23	'bicycle')	22	'bicycle')

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/coco.py+0-1

	@@ -21,7 +21,6 @@ from .registry import DATASETS
21		21
22	@DATASETS.register_module	22	@DATASETS.register_module
23	class CocoDataset(CustomDataset):	23	class CocoDataset(CustomDataset):
24	-
25	CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',	24	CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
26	'train', 'truck', 'boat', 'traffic_light', 'fire_hydrant',	25	'train', 'truck', 'boat', 'traffic_light', 'fire_hydrant',
27	'stop_sign', 'parking_meter', 'bench', 'bird', 'cat', 'dog',	26	'stop_sign', 'parking_meter', 'bench', 'bird', 'cat', 'dog',

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/loader/build_loader.py+1-0

	@@ -24,6 +24,7 @@ from .sampler import DistributedGroupSampler, DistributedSampler, GroupSampler
24	if platform.system() != 'Windows':	24	if platform.system() != 'Windows':
25	# https://github.com/pytorch/pytorch/issues/973	25	# https://github.com/pytorch/pytorch/issues/973
26	import resource	26	import resource
		27	+
27	rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)	28	rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
28	resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1]))	29	resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1]))
29		30

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/pipelines/transforms.py+8-8

	@@ -317,7 +317,7 @@ class Pad(object):
317	if padded_masks:	317	if padded_masks:
318	results[key] = np.stack(padded_masks, axis=0)	318	results[key] = np.stack(padded_masks, axis=0)
319	else:	319	else:
320	- results[key] = np.empty((0, ) + pad_shape, dtype=np.uint8)	320	+ results[key] = np.empty((0,) + pad_shape, dtype=np.uint8)
321		321
322	def _pad_seg(self, results):	322	def _pad_seg(self, results):
323	for key in results.get('seg_fields', []):	323	for key in results.get('seg_fields', []):
	@@ -409,7 +409,7 @@ class RandomCrop(object):
409	if 'gt_bboxes' in results:	409	if 'gt_bboxes' in results:
410	gt_bboxes = results['gt_bboxes']	410	gt_bboxes = results['gt_bboxes']
411	valid_inds = (gt_bboxes[:, 2] > gt_bboxes[:, 0]) & (	411	valid_inds = (gt_bboxes[:, 2] > gt_bboxes[:, 0]) & (
412	- gt_bboxes[:, 3] > gt_bboxes[:, 1])	412	+ gt_bboxes[:, 3] > gt_bboxes[:, 1])
413	# if no gt bbox remains after cropping, just skip this image	413	# if no gt bbox remains after cropping, just skip this image
414	if not np.any(valid_inds):	414	if not np.any(valid_inds):
415	return None	415	return None
	@@ -422,7 +422,7 @@ class RandomCrop(object):
422	valid_gt_masks = []	422	valid_gt_masks = []
423	for i in np.where(valid_inds)[0]:	423	for i in np.where(valid_inds)[0]:
424	gt_mask = results['gt_masks'][i][crop_y1:crop_y2,	424	gt_mask = results['gt_masks'][i][crop_y1:crop_y2,
425	- crop_x1:crop_x2]	425	+ crop_x1:crop_x2]
426	valid_gt_masks.append(gt_mask)	426	valid_gt_masks.append(gt_mask)
427	results['gt_masks'] = np.stack(valid_gt_masks)	427	results['gt_masks'] = np.stack(valid_gt_masks)
428		428
	@@ -540,8 +540,8 @@ class PhotoMetricDistortion(object):
540	repr_str = self.__class__.__name__	540	repr_str = self.__class__.__name__
541	repr_str += ('(brightness_delta={}, contrast_range={}, '	541	repr_str += ('(brightness_delta={}, contrast_range={}, '
542	'saturation_range={}, hue_delta={})').format(	542	'saturation_range={}, hue_delta={})').format(
543	- self.brightness_delta, self.contrast_range,	543	+ self.brightness_delta, self.contrast_range,
544	- self.saturation_range, self.hue_delta)	544	+ self.saturation_range, self.hue_delta)
545	return repr_str	545	return repr_str
546		546
547		547
	@@ -616,8 +616,8 @@ class Expand(object):
616	repr_str = self.__class__.__name__	616	repr_str = self.__class__.__name__
617	repr_str += '(mean={}, to_rgb={}, ratio_range={}, ' \	617	repr_str += '(mean={}, to_rgb={}, ratio_range={}, ' \
618	'seg_ignore_label={})'.format(	618	'seg_ignore_label={})'.format(
619	- self.mean, self.to_rgb, self.ratio_range,	619	+ self.mean, self.to_rgb, self.ratio_range,
620	- self.seg_ignore_label)	620	+ self.seg_ignore_label)
621	return repr_str	621	return repr_str
622		622
623		623
	@@ -700,7 +700,7 @@ class MinIoURandomCrop(object):
700	# not tested	700	# not tested
701	if 'gt_semantic_seg' in results:	701	if 'gt_semantic_seg' in results:
702	results['gt_semantic_seg'] = results['gt_semantic_seg'][	702	results['gt_semantic_seg'] = results['gt_semantic_seg'][
703	- patch[1]:patch[3], patch[0]:patch[2]]	703	+ patch[1]:patch[3], patch[0]:patch[2]]
704	return results	704	return results
705		705
706	def __repr__(self):	706	def __repr__(self):

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/voc.py+0-1

	@@ -18,7 +18,6 @@ from .xml_style import XMLDataset
18		18
19	@DATASETS.register_module	19	@DATASETS.register_module
20	class VOCDataset(XMLDataset):	20	class VOCDataset(XMLDataset):
21	-
22	CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',	21	CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',
23	'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',	22	'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
24	'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',	23	'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/wider_face.py+1-1

	@@ -28,7 +28,7 @@ class WIDERFaceDataset(XMLDataset):
28	Conversion scripts can be found in	28	Conversion scripts can be found in
29	https://github.com/sovrasov/wider-face-pascal-voc-annotations	29	https://github.com/sovrasov/wider-face-pascal-voc-annotations
30	"""	30	"""
31	- CLASSES = ('face', )	31	+ CLASSES = ('face',)
32		32
33	def __init__(self, **kwargs):	33	def __init__(self, **kwargs):
34	super(WIDERFaceDataset, self).__init__(**kwargs)	34	super(WIDERFaceDataset, self).__init__(**kwargs)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/datasets/xml_style.py+2-2

	@@ -82,13 +82,13 @@ class XMLDataset(CustomDataset):
82	labels.append(label)	82	labels.append(label)
83	if not bboxes:	83	if not bboxes:
84	bboxes = np.zeros((0, 4))	84	bboxes = np.zeros((0, 4))
85	- labels = np.zeros((0, ))	85	+ labels = np.zeros((0,))
86	else:	86	else:
87	bboxes = np.array(bboxes, ndmin=2) - 1	87	bboxes = np.array(bboxes, ndmin=2) - 1
88	labels = np.array(labels)	88	labels = np.array(labels)
89	if not bboxes_ignore:	89	if not bboxes_ignore:
90	bboxes_ignore = np.zeros((0, 4))	90	bboxes_ignore = np.zeros((0, 4))
91	- labels_ignore = np.zeros((0, ))	91	+ labels_ignore = np.zeros((0,))
92	else:	92	else:
93	bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1	93	bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1
94	labels_ignore = np.array(labels_ignore)	94	labels_ignore = np.array(labels_ignore)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/__init__.py+1-0

	@@ -23,6 +23,7 @@ from .mask_heads import * # noqa: F401,F403
23	from .necks import * # noqa: F401,F403	23	from .necks import * # noqa: F401,F403
24	from .registry import (BACKBONES, DETECTORS, HEADS, LOSSES, NECKS,	24	from .registry import (BACKBONES, DETECTORS, HEADS, LOSSES, NECKS,
25	ROI_EXTRACTORS, SHARED_HEADS)	25	ROI_EXTRACTORS, SHARED_HEADS)
		26	+
26	# from .roi_extractors import * # noqa: F401,F403	27	# from .roi_extractors import * # noqa: F401,F403
27	# from .shared_heads import * # noqa: F401,F403	28	# from .shared_heads import * # noqa: F401,F403
28		29

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/__init__.py+1-0

	@@ -27,6 +27,7 @@
27	# from .ssd_head import SSDHead	27	# from .ssd_head import SSDHead
28	from .solo_head import SOLOHead	28	from .solo_head import SOLOHead
29	from .solov2_head import SOLOv2Head	29	from .solov2_head import SOLOv2Head
		30	+
30	# from .solov2_light_head import SOLOv2LightHead	31	# from .solov2_light_head import SOLOv2LightHead
31	# from .decoupled_solo_head import DecoupledSOLOHead	32	# from .decoupled_solo_head import DecoupledSOLOHead
32	# from .decoupled_solo_light_head import DecoupledSOLOLightHead	33	# from .decoupled_solo_light_head import DecoupledSOLOLightHead

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/atss_head.py+25-25

	@@ -67,7 +67,7 @@ class ATSSHead(AnchorHead):
67	self.norm_cfg = norm_cfg	67	self.norm_cfg = norm_cfg
68		68
69	octave_scales = np.array(	69	octave_scales = np.array(
70	- [2**(i / scales_per_octave) for i in range(scales_per_octave)])	70	+ [2 ** (i / scales_per_octave) for i in range(scales_per_octave)])
71	anchor_scales = octave_scales * octave_base_scale	71	anchor_scales = octave_scales * octave_base_scale
72	super(ATSSHead, self).__init__(	72	super(ATSSHead, self).__init__(
73	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)	73	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)
	@@ -226,18 +226,18 @@ class ATSSHead(AnchorHead):
226	torch.tensor(num_total_pos).cuda()).item()	226	torch.tensor(num_total_pos).cuda()).item()
227	num_total_samples = max(num_total_samples, 1.0)	227	num_total_samples = max(num_total_samples, 1.0)
228		228
229	- losses_cls, losses_bbox, loss_centerness,\	229	+ losses_cls, losses_bbox, loss_centerness, \
230	- bbox_avg_factor = multi_apply(	230	+ bbox_avg_factor = multi_apply(
231	- self.loss_single,	231	+ self.loss_single,
232	- anchor_list,	232	+ anchor_list,
233	- cls_scores,	233	+ cls_scores,
234	- bbox_preds,	234	+ bbox_preds,
235	- centernesses,	235	+ centernesses,
236	- labels_list,	236	+ labels_list,
237	- label_weights_list,	237	+ label_weights_list,
238	- bbox_targets_list,	238	+ bbox_targets_list,
239	- num_total_samples=num_total_samples,	239	+ num_total_samples=num_total_samples,
240	- cfg=cfg)	240	+ cfg=cfg)
241		241
242	bbox_avg_factor = sum(bbox_avg_factor)	242	bbox_avg_factor = sum(bbox_avg_factor)
243	bbox_avg_factor = reduce_mean(bbox_avg_factor).item()	243	bbox_avg_factor = reduce_mean(bbox_avg_factor).item()
	@@ -394,17 +394,17 @@ class ATSSHead(AnchorHead):
394	gt_labels_list = [None for _ in range(num_imgs)]	394	gt_labels_list = [None for _ in range(num_imgs)]
395	(all_anchors, all_labels, all_label_weights, all_bbox_targets,	395	(all_anchors, all_labels, all_label_weights, all_bbox_targets,
396	all_bbox_weights, pos_inds_list, neg_inds_list) = multi_apply(	396	all_bbox_weights, pos_inds_list, neg_inds_list) = multi_apply(
397	- self.atss_target_single,	397	+ self.atss_target_single,
398	- anchor_list,	398	+ anchor_list,
399	- valid_flag_list,	399	+ valid_flag_list,
400	- num_level_anchors_list,	400	+ num_level_anchors_list,
401	- gt_bboxes_list,	401	+ gt_bboxes_list,
402	- gt_bboxes_ignore_list,	402	+ gt_bboxes_ignore_list,
403	- gt_labels_list,	403	+ gt_labels_list,
404	- img_metas,	404	+ img_metas,
405	- cfg=cfg,	405	+ cfg=cfg,
406	- label_channels=label_channels,	406	+ label_channels=label_channels,
407	- unmap_outputs=unmap_outputs)	407	+ unmap_outputs=unmap_outputs)
408	# no valid anchors	408	# no valid anchors
409	if any([labels is None for labels in all_labels]):	409	if any([labels is None for labels in all_labels]):
410	return None	410	return None
	@@ -439,7 +439,7 @@ class ATSSHead(AnchorHead):
439	img_meta['img_shape'][:2],	439	img_meta['img_shape'][:2],
440	cfg.allowed_border)	440	cfg.allowed_border)
441	if not inside_flags.any():	441	if not inside_flags.any():
442	- return (None, ) * 6	442	+ return (None,) * 6
443	# assign gt and sample anchors	443	# assign gt and sample anchors
444	anchors = flat_anchors[inside_flags, :]	444	anchors = flat_anchors[inside_flags, :]
445		445

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/decoupled_solo_head.py+41-35

	@@ -25,6 +25,7 @@ from ..utils import bias_init_with_prob, ConvModule
25		25
26	INF = 1e8	26	INF = 1e8
27		27
		28	+
28	def center_of_mass(bitmasks):	29	def center_of_mass(bitmasks):
29	_, h, w = bitmasks.size()	30	_, h, w = bitmasks.size()
30	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)	31	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)
	@@ -37,6 +38,7 @@ def center_of_mass(bitmasks):
37	center_y = m01 / m00	38	center_y = m01 / m00
38	return center_x, center_y	39	return center_x, center_y
39		40
		41	+
40	def points_nms(heat, kernel=2):	42	def points_nms(heat, kernel=2):
41	# kernel must be 2	43	# kernel must be 2
42	hmax = nn.functional.max_pool2d(	44	hmax = nn.functional.max_pool2d(
	@@ -44,6 +46,7 @@ def points_nms(heat, kernel=2):
44	keep = (hmax[:, :, :-1, :-1] == heat).float()	46	keep = (hmax[:, :, :-1, :-1] == heat).float()
45	return heat * keep	47	return heat * keep
46		48
		49	+
47	def dice_loss(input, target):	50	def dice_loss(input, target):
48	input = input.contiguous().view(input.size()[0], -1)	51	input = input.contiguous().view(input.size()[0], -1)
49	target = target.contiguous().view(target.size()[0], -1).float()	52	target = target.contiguous().view(target.size()[0], -1).float()
	@@ -52,7 +55,8 @@ def dice_loss(input, target):
52	b = torch.sum(input * input, 1) + 0.001	55	b = torch.sum(input * input, 1) + 0.001
53	c = torch.sum(target * target, 1) + 0.001	56	c = torch.sum(target * target, 1) + 0.001
54	d = (2 * a) / (b + c)	57	d = (2 * a) / (b + c)
55	- return 1-d	58	+ return 1 - d
		59	+
56		60
57	@HEADS.register_module	61	@HEADS.register_module
58	class DecoupledSOLOHead(nn.Module):	62	class DecoupledSOLOHead(nn.Module):
	@@ -166,10 +170,10 @@ class DecoupledSOLOHead(nn.Module):
166	return ins_pred_x, ins_pred_y, cate_pred	170	return ins_pred_x, ins_pred_y, cate_pred
167		171
168	def split_feats(self, feats):	172	def split_feats(self, feats):
169	- return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),	173	+ return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),
170	- feats[1],	174	+ feats[1],
171	- feats[2],	175	+ feats[2],
172	- feats[3],	176	+ feats[3],
173	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))	177	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))
174		178
175	def forward_single(self, x, idx, eval=False, upsampled_size=None):	179	def forward_single(self, x, idx, eval=False, upsampled_size=None):
	@@ -198,7 +202,7 @@ class DecoupledSOLOHead(nn.Module):
198	# cate branch	202	# cate branch
199	for i, cate_layer in enumerate(self.cate_convs):	203	for i, cate_layer in enumerate(self.cate_convs):
200	if i == self.cate_down_pos:	204	if i == self.cate_down_pos:
201	- seg_num_grid = self.seg_num_grids[idx]	205	+ seg_num_grid = self.seg_num_grids[idx]
202	cate_feat = F.interpolate(cate_feat, size=seg_num_grid, mode='bilinear')	206	cate_feat = F.interpolate(cate_feat, size=seg_num_grid, mode='bilinear')
203	cate_feat = cate_layer(cate_feat)	207	cate_feat = cate_layer(cate_feat)
204		208
	@@ -236,16 +240,16 @@ class DecoupledSOLOHead(nn.Module):
236	for ins_labels_level, ins_ind_labels_level in zip(zip(ins_label_list), zip(ins_ind_label_list))]	240	for ins_labels_level, ins_ind_labels_level in zip(zip(ins_label_list), zip(ins_ind_label_list))]
237		241
238	ins_preds_x_final = [torch.cat([ins_preds_level_img_x[ins_ind_labels_level_img[:, 1], ...]	242	ins_preds_x_final = [torch.cat([ins_preds_level_img_x[ins_ind_labels_level_img[:, 1], ...]
239	- for ins_preds_level_img_x, ins_ind_labels_level_img in	243	+ for ins_preds_level_img_x, ins_ind_labels_level_img in
240	- zip(ins_preds_level_x, ins_ind_labels_level)], 0)	244	+ zip(ins_preds_level_x, ins_ind_labels_level)], 0)
241	- for ins_preds_level_x, ins_ind_labels_level in	245	+ for ins_preds_level_x, ins_ind_labels_level in
242	- zip(ins_preds_x, zip(*ins_ind_label_list_xy))]	246	+ zip(ins_preds_x, zip(*ins_ind_label_list_xy))]
243		247
244	ins_preds_y_final = [torch.cat([ins_preds_level_img_y[ins_ind_labels_level_img[:, 0], ...]	248	ins_preds_y_final = [torch.cat([ins_preds_level_img_y[ins_ind_labels_level_img[:, 0], ...]
245	- for ins_preds_level_img_y, ins_ind_labels_level_img in	249	+ for ins_preds_level_img_y, ins_ind_labels_level_img in
246	- zip(ins_preds_level_y, ins_ind_labels_level)], 0)	250	+ zip(ins_preds_level_y, ins_ind_labels_level)], 0)
247	- for ins_preds_level_y, ins_ind_labels_level in	251	+ for ins_preds_level_y, ins_ind_labels_level in
248	- zip(ins_preds_y, zip(*ins_ind_label_list_xy))]	252	+ zip(ins_preds_y, zip(*ins_ind_label_list_xy))]
249		253
250	num_ins = 0.	254	num_ins = 0.
251	# dice loss	255	# dice loss
	@@ -255,7 +259,7 @@ class DecoupledSOLOHead(nn.Module):
255	if mask_n == 0:	259	if mask_n == 0:
256	continue	260	continue
257	num_ins += mask_n	261	num_ins += mask_n
258	- input = (input_x.sigmoid())*(input_y.sigmoid())	262	+ input = (input_x.sigmoid()) * (input_y.sigmoid())
259	loss_ins.append(dice_loss(input, target))	263	loss_ins.append(dice_loss(input, target))
260		264
261	loss_ins = torch.cat(loss_ins).mean() * self.ins_loss_weight	265	loss_ins = torch.cat(loss_ins).mean() * self.ins_loss_weight
	@@ -280,10 +284,10 @@ class DecoupledSOLOHead(nn.Module):
280	loss_cate=loss_cate)	284	loss_cate=loss_cate)
281		285
282	def solo_target_single(self,	286	def solo_target_single(self,
283	- gt_bboxes_raw,	287	+ gt_bboxes_raw,
284	- gt_labels_raw,	288	+ gt_labels_raw,
285	- gt_masks_raw,	289	+ gt_masks_raw,
286	- featmap_sizes=None):	290	+ featmap_sizes=None):
287		291
288	device = gt_labels_raw[0].device	292	device = gt_labels_raw[0].device
289	# ins	293	# ins
	@@ -296,9 +300,9 @@ class DecoupledSOLOHead(nn.Module):
296	for (lower_bound, upper_bound), stride, featmap_size, num_grid \	300	for (lower_bound, upper_bound), stride, featmap_size, num_grid \
297	in zip(self.scale_ranges, self.strides, featmap_sizes, self.seg_num_grids):	301	in zip(self.scale_ranges, self.strides, featmap_sizes, self.seg_num_grids):
298		302
299	- ins_label = torch.zeros([num_grid**2, featmap_size[0], featmap_size[1]], dtype=torch.uint8, device=device)	303	+ ins_label = torch.zeros([num_grid ** 2, featmap_size[0], featmap_size[1]], dtype=torch.uint8, device=device)
300	cate_label = torch.zeros([num_grid, num_grid], dtype=torch.int64, device=device)	304	cate_label = torch.zeros([num_grid, num_grid], dtype=torch.int64, device=device)
301	- ins_ind_label = torch.zeros([num_grid**2], dtype=torch.bool, device=device)	305	+ ins_ind_label = torch.zeros([num_grid ** 2], dtype=torch.bool, device=device)
302		306
303	hit_indices = ((gt_areas >= lower_bound) & (gt_areas <= upper_bound)).nonzero().flatten()	307	hit_indices = ((gt_areas >= lower_bound) & (gt_areas <= upper_bound)).nonzero().flatten()
304		308
	@@ -324,9 +328,12 @@ class DecoupledSOLOHead(nn.Module):
324	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0	328	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0
325		329
326	output_stride = stride / 2	330	output_stride = stride / 2
327	- for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels, half_hs, half_ws, center_hs, center_ws, valid_mask_flags):	331	+ for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels,
		332	+ half_hs, half_ws,
		333	+ center_hs, center_ws,
		334	+ valid_mask_flags):
328	if not valid_mask_flag:	335	if not valid_mask_flag:
329	- continue	336	+ continue
330	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)	337	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)
331	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))	338	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))
332	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))	339	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))
	@@ -337,18 +344,18 @@ class DecoupledSOLOHead(nn.Module):
337	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))	344	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))
338	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))	345	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))
339		346
340	- top = max(top_box, coord_h-1)	347	+ top = max(top_box, coord_h - 1)
341	- down = min(down_box, coord_h+1)	348	+ down = min(down_box, coord_h + 1)
342	- left = max(coord_w-1, left_box)	349	+ left = max(coord_w - 1, left_box)
343	- right = min(right_box, coord_w+1)	350	+ right = min(right_box, coord_w + 1)
344		351
345	# squared	352	# squared
346	- cate_label[top:(down+1), left:(right+1)] = gt_label	353	+ cate_label[top:(down + 1), left:(right + 1)] = gt_label
347	# ins	354	# ins
348	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)	355	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)
349	seg_mask = torch.from_numpy(seg_mask).to(device=device)	356	seg_mask = torch.from_numpy(seg_mask).to(device=device)
350	- for i in range(top, down+1):	357	+ for i in range(top, down + 1):
351	- for j in range(left, right+1):	358	+ for j in range(left, right + 1):
352	label = int(i * num_grid + j)	359	label = int(i * num_grid + j)
353	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask	360	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask
354	ins_ind_label[label] = True	361	ins_ind_label[label] = True
	@@ -404,7 +411,6 @@ class DecoupledSOLOHead(nn.Module):
404	cfg,	411	cfg,
405	rescale=False, debug=False):	412	rescale=False, debug=False):
406		413
407	-
408	# overall info.	414	# overall info.
409	h, w, _ = img_shape	415	h, w, _ = img_shape
410	upsampled_size_out = (featmap_size[0] * 4, featmap_size[1] * 4)	416	upsampled_size_out = (featmap_size[0] * 4, featmap_size[1] * 4)
	@@ -489,10 +495,10 @@ class DecoupledSOLOHead(nn.Module):
489	cate_labels = cate_labels[sort_inds]	495	cate_labels = cate_labels[sort_inds]
490		496
491	seg_masks_soft = F.interpolate(seg_masks_soft.unsqueeze(0),	497	seg_masks_soft = F.interpolate(seg_masks_soft.unsqueeze(0),
492	- size=upsampled_size_out,	498	+ size=upsampled_size_out,
493	- mode='bilinear')[:, :, :h, :w]	499	+ mode='bilinear')[:, :, :h, :w]
494	seg_masks = F.interpolate(seg_masks_soft,	500	seg_masks = F.interpolate(seg_masks_soft,
495	- size=ori_shape[:2],	501	+ size=ori_shape[:2],
496	- mode='bilinear').squeeze(0)	502	+ mode='bilinear').squeeze(0)
497	seg_masks = seg_masks > cfg.mask_thr	503	seg_masks = seg_masks > cfg.mask_thr
498	return seg_masks, cate_labels, cate_scores	504	return seg_masks, cate_labels, cate_scores

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/decoupled_solo_light_head.py+41-35

	@@ -25,6 +25,7 @@ from ..utils import bias_init_with_prob, ConvModule
25		25
26	INF = 1e8	26	INF = 1e8
27		27
		28	+
28	def center_of_mass(bitmasks):	29	def center_of_mass(bitmasks):
29	_, h, w = bitmasks.size()	30	_, h, w = bitmasks.size()
30	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)	31	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)
	@@ -37,6 +38,7 @@ def center_of_mass(bitmasks):
37	center_y = m01 / m00	38	center_y = m01 / m00
38	return center_x, center_y	39	return center_x, center_y
39		40
		41	+
40	def points_nms(heat, kernel=2):	42	def points_nms(heat, kernel=2):
41	# kernel must be 2	43	# kernel must be 2
42	hmax = nn.functional.max_pool2d(	44	hmax = nn.functional.max_pool2d(
	@@ -44,6 +46,7 @@ def points_nms(heat, kernel=2):
44	keep = (hmax[:, :, :-1, :-1] == heat).float()	46	keep = (hmax[:, :, :-1, :-1] == heat).float()
45	return heat * keep	47	return heat * keep
46		48
		49	+
47	def dice_loss(input, target):	50	def dice_loss(input, target):
48	input = input.contiguous().view(input.size()[0], -1)	51	input = input.contiguous().view(input.size()[0], -1)
49	target = target.contiguous().view(target.size()[0], -1).float()	52	target = target.contiguous().view(target.size()[0], -1).float()
	@@ -52,7 +55,8 @@ def dice_loss(input, target):
52	b = torch.sum(input * input, 1) + 0.001	55	b = torch.sum(input * input, 1) + 0.001
53	c = torch.sum(target * target, 1) + 0.001	56	c = torch.sum(target * target, 1) + 0.001
54	d = (2 * a) / (b + c)	57	d = (2 * a) / (b + c)
55	- return 1-d	58	+ return 1 - d
		59	+
56		60
57	@HEADS.register_module	61	@HEADS.register_module
58	class DecoupledSOLOLightHead(nn.Module):	62	class DecoupledSOLOLightHead(nn.Module):
	@@ -163,10 +167,10 @@ class DecoupledSOLOLightHead(nn.Module):
163	return ins_pred_x, ins_pred_y, cate_pred	167	return ins_pred_x, ins_pred_y, cate_pred
164		168
165	def split_feats(self, feats):	169	def split_feats(self, feats):
166	- return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),	170	+ return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),
167	- feats[1],	171	+ feats[1],
168	- feats[2],	172	+ feats[2],
169	- feats[3],	173	+ feats[3],
170	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))	174	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))
171		175
172	def forward_single(self, x, idx, eval=False, upsampled_size=None):	176	def forward_single(self, x, idx, eval=False, upsampled_size=None):
	@@ -193,7 +197,7 @@ class DecoupledSOLOLightHead(nn.Module):
193	# cate branch	197	# cate branch
194	for i, cate_layer in enumerate(self.cate_convs):	198	for i, cate_layer in enumerate(self.cate_convs):
195	if i == self.cate_down_pos:	199	if i == self.cate_down_pos:
196	- seg_num_grid = self.seg_num_grids[idx]	200	+ seg_num_grid = self.seg_num_grids[idx]
197	cate_feat = F.interpolate(cate_feat, size=seg_num_grid, mode='bilinear')	201	cate_feat = F.interpolate(cate_feat, size=seg_num_grid, mode='bilinear')
198	cate_feat = cate_layer(cate_feat)	202	cate_feat = cate_layer(cate_feat)
199		203
	@@ -231,16 +235,16 @@ class DecoupledSOLOLightHead(nn.Module):
231	for ins_labels_level, ins_ind_labels_level in zip(zip(ins_label_list), zip(ins_ind_label_list))]	235	for ins_labels_level, ins_ind_labels_level in zip(zip(ins_label_list), zip(ins_ind_label_list))]
232		236
233	ins_preds_x_final = [torch.cat([ins_preds_level_img_x[ins_ind_labels_level_img[:, 1], ...]	237	ins_preds_x_final = [torch.cat([ins_preds_level_img_x[ins_ind_labels_level_img[:, 1], ...]
234	- for ins_preds_level_img_x, ins_ind_labels_level_img in	238	+ for ins_preds_level_img_x, ins_ind_labels_level_img in
235	- zip(ins_preds_level_x, ins_ind_labels_level)], 0)	239	+ zip(ins_preds_level_x, ins_ind_labels_level)], 0)
236	- for ins_preds_level_x, ins_ind_labels_level in	240	+ for ins_preds_level_x, ins_ind_labels_level in
237	- zip(ins_preds_x, zip(*ins_ind_label_list_xy))]	241	+ zip(ins_preds_x, zip(*ins_ind_label_list_xy))]
238		242
239	ins_preds_y_final = [torch.cat([ins_preds_level_img_y[ins_ind_labels_level_img[:, 0], ...]	243	ins_preds_y_final = [torch.cat([ins_preds_level_img_y[ins_ind_labels_level_img[:, 0], ...]
240	- for ins_preds_level_img_y, ins_ind_labels_level_img in	244	+ for ins_preds_level_img_y, ins_ind_labels_level_img in
241	- zip(ins_preds_level_y, ins_ind_labels_level)], 0)	245	+ zip(ins_preds_level_y, ins_ind_labels_level)], 0)
242	- for ins_preds_level_y, ins_ind_labels_level in	246	+ for ins_preds_level_y, ins_ind_labels_level in
243	- zip(ins_preds_y, zip(*ins_ind_label_list_xy))]	247	+ zip(ins_preds_y, zip(*ins_ind_label_list_xy))]
244		248
245	num_ins = 0.	249	num_ins = 0.
246	# dice loss	250	# dice loss
	@@ -250,7 +254,7 @@ class DecoupledSOLOLightHead(nn.Module):
250	if mask_n == 0:	254	if mask_n == 0:
251	continue	255	continue
252	num_ins += mask_n	256	num_ins += mask_n
253	- input = (input_x.sigmoid())*(input_y.sigmoid())	257	+ input = (input_x.sigmoid()) * (input_y.sigmoid())
254	loss_ins.append(dice_loss(input, target))	258	loss_ins.append(dice_loss(input, target))
255		259
256	loss_ins = torch.cat(loss_ins).mean() * self.ins_loss_weight	260	loss_ins = torch.cat(loss_ins).mean() * self.ins_loss_weight
	@@ -275,10 +279,10 @@ class DecoupledSOLOLightHead(nn.Module):
275	loss_cate=loss_cate)	279	loss_cate=loss_cate)
276		280
277	def solo_target_single(self,	281	def solo_target_single(self,
278	- gt_bboxes_raw,	282	+ gt_bboxes_raw,
279	- gt_labels_raw,	283	+ gt_labels_raw,
280	- gt_masks_raw,	284	+ gt_masks_raw,
281	- featmap_sizes=None):	285	+ featmap_sizes=None):
282		286
283	device = gt_labels_raw[0].device	287	device = gt_labels_raw[0].device
284	# ins	288	# ins
	@@ -291,9 +295,9 @@ class DecoupledSOLOLightHead(nn.Module):
291	for (lower_bound, upper_bound), stride, featmap_size, num_grid \	295	for (lower_bound, upper_bound), stride, featmap_size, num_grid \
292	in zip(self.scale_ranges, self.strides, featmap_sizes, self.seg_num_grids):	296	in zip(self.scale_ranges, self.strides, featmap_sizes, self.seg_num_grids):
293		297
294	- ins_label = torch.zeros([num_grid**2, featmap_size[0], featmap_size[1]], dtype=torch.uint8, device=device)	298	+ ins_label = torch.zeros([num_grid ** 2, featmap_size[0], featmap_size[1]], dtype=torch.uint8, device=device)
295	cate_label = torch.zeros([num_grid, num_grid], dtype=torch.int64, device=device)	299	cate_label = torch.zeros([num_grid, num_grid], dtype=torch.int64, device=device)
296	- ins_ind_label = torch.zeros([num_grid**2], dtype=torch.bool, device=device)	300	+ ins_ind_label = torch.zeros([num_grid ** 2], dtype=torch.bool, device=device)
297		301
298	hit_indices = ((gt_areas >= lower_bound) & (gt_areas <= upper_bound)).nonzero().flatten()	302	hit_indices = ((gt_areas >= lower_bound) & (gt_areas <= upper_bound)).nonzero().flatten()
299		303
	@@ -319,9 +323,12 @@ class DecoupledSOLOLightHead(nn.Module):
319	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0	323	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0
320		324
321	output_stride = stride / 2	325	output_stride = stride / 2
322	- for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels, half_hs, half_ws, center_hs, center_ws, valid_mask_flags):	326	+ for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels,
		327	+ half_hs, half_ws,
		328	+ center_hs, center_ws,
		329	+ valid_mask_flags):
323	if not valid_mask_flag:	330	if not valid_mask_flag:
324	- continue	331	+ continue
325	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)	332	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)
326	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))	333	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))
327	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))	334	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))
	@@ -332,18 +339,18 @@ class DecoupledSOLOLightHead(nn.Module):
332	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))	339	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))
333	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))	340	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))
334		341
335	- top = max(top_box, coord_h-1)	342	+ top = max(top_box, coord_h - 1)
336	- down = min(down_box, coord_h+1)	343	+ down = min(down_box, coord_h + 1)
337	- left = max(coord_w-1, left_box)	344	+ left = max(coord_w - 1, left_box)
338	- right = min(right_box, coord_w+1)	345	+ right = min(right_box, coord_w + 1)
339		346
340	# squared	347	# squared
341	- cate_label[top:(down+1), left:(right+1)] = gt_label	348	+ cate_label[top:(down + 1), left:(right + 1)] = gt_label
342	# ins	349	# ins
343	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)	350	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)
344	seg_mask = torch.from_numpy(seg_mask).to(device=device)	351	seg_mask = torch.from_numpy(seg_mask).to(device=device)
345	- for i in range(top, down+1):	352	+ for i in range(top, down + 1):
346	- for j in range(left, right+1):	353	+ for j in range(left, right + 1):
347	label = int(i * num_grid + j)	354	label = int(i * num_grid + j)
348	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask	355	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask
349	ins_ind_label[label] = True	356	ins_ind_label[label] = True
	@@ -399,7 +406,6 @@ class DecoupledSOLOLightHead(nn.Module):
399	cfg,	406	cfg,
400	rescale=False, debug=False):	407	rescale=False, debug=False):
401		408
402	-
403	# overall info.	409	# overall info.
404	h, w, _ = img_shape	410	h, w, _ = img_shape
405	upsampled_size_out = (featmap_size[0] * 4, featmap_size[1] * 4)	411	upsampled_size_out = (featmap_size[0] * 4, featmap_size[1] * 4)
	@@ -484,10 +490,10 @@ class DecoupledSOLOLightHead(nn.Module):
484	cate_labels = cate_labels[sort_inds]	490	cate_labels = cate_labels[sort_inds]
485		491
486	seg_masks_soft = F.interpolate(seg_masks_soft.unsqueeze(0),	492	seg_masks_soft = F.interpolate(seg_masks_soft.unsqueeze(0),
487	- size=upsampled_size_out,	493	+ size=upsampled_size_out,
488	- mode='bilinear')[:, :, :h, :w]	494	+ mode='bilinear')[:, :, :h, :w]
489	seg_masks = F.interpolate(seg_masks_soft,	495	seg_masks = F.interpolate(seg_masks_soft,
490	- size=ori_shape[:2],	496	+ size=ori_shape[:2],
491	- mode='bilinear').squeeze(0)	497	+ mode='bilinear').squeeze(0)
492	seg_masks = seg_masks > cfg.mask_thr	498	seg_masks = seg_masks > cfg.mask_thr
493	return seg_masks, cate_labels, cate_scores	499	return seg_masks, cate_labels, cate_scores

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/fcos_head.py+5-5

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/fovea_head.py+8-8

	@@ -297,16 +297,16 @@ class FoveaHead(nn.Module):
297	half_h = 0.5 * (gt_bboxes[:, 3] - gt_bboxes[:, 1])	297	half_h = 0.5 * (gt_bboxes[:, 3] - gt_bboxes[:, 1])
298	# valid fovea area: left, right, top, down	298	# valid fovea area: left, right, top, down
299	pos_left = torch.ceil(	299	pos_left = torch.ceil(
300	- gt_bboxes[:, 0] + (1 - self.sigma) * half_w - 0.5).long().\	300	+ gt_bboxes[:, 0] + (1 - self.sigma) * half_w - 0.5).long(). \
301	clamp(0, featmap_size[1] - 1)	301	clamp(0, featmap_size[1] - 1)
302	pos_right = torch.floor(	302	pos_right = torch.floor(
303	- gt_bboxes[:, 0] + (1 + self.sigma) * half_w - 0.5).long().\	303	+ gt_bboxes[:, 0] + (1 + self.sigma) * half_w - 0.5).long(). \
304	clamp(0, featmap_size[1] - 1)	304	clamp(0, featmap_size[1] - 1)
305	pos_top = torch.ceil(	305	pos_top = torch.ceil(
306	- gt_bboxes[:, 1] + (1 - self.sigma) * half_h - 0.5).long().\	306	+ gt_bboxes[:, 1] + (1 - self.sigma) * half_h - 0.5).long(). \
307	clamp(0, featmap_size[0] - 1)	307	clamp(0, featmap_size[0] - 1)
308	pos_down = torch.floor(	308	pos_down = torch.floor(
309	- gt_bboxes[:, 1] + (1 + self.sigma) * half_h - 0.5).long().\	309	+ gt_bboxes[:, 1] + (1 + self.sigma) * half_h - 0.5).long(). \
310	clamp(0, featmap_size[0] - 1)	310	clamp(0, featmap_size[0] - 1)
311	for px1, py1, px2, py2, label, (gt_x1, gt_y1, gt_x2, gt_y2) in \	311	for px1, py1, px2, py2, label, (gt_x1, gt_y1, gt_x2, gt_y2) in \
312	zip(pos_left, pos_top, pos_right, pos_down, gt_labels,	312	zip(pos_left, pos_top, pos_right, pos_down, gt_labels,
	@@ -378,13 +378,13 @@ class FoveaHead(nn.Module):
378	scores = scores[topk_inds, :]	378	scores = scores[topk_inds, :]
379	y = y[topk_inds]	379	y = y[topk_inds]
380	x = x[topk_inds]	380	x = x[topk_inds]
381	- x1 = (stride * x - base_len * bbox_pred[:, 0]).\	381	+ x1 = (stride * x - base_len * bbox_pred[:, 0]). \
382	clamp(min=0, max=img_shape[1] - 1)	382	clamp(min=0, max=img_shape[1] - 1)
383	- y1 = (stride * y - base_len * bbox_pred[:, 1]).\	383	+ y1 = (stride * y - base_len * bbox_pred[:, 1]). \
384	clamp(min=0, max=img_shape[0] - 1)	384	clamp(min=0, max=img_shape[0] - 1)
385	- x2 = (stride * x + base_len * bbox_pred[:, 2]).\	385	+ x2 = (stride * x + base_len * bbox_pred[:, 2]). \
386	clamp(min=0, max=img_shape[1] - 1)	386	clamp(min=0, max=img_shape[1] - 1)
387	- y2 = (stride * y + base_len * bbox_pred[:, 3]).\	387	+ y2 = (stride * y + base_len * bbox_pred[:, 3]). \
388	clamp(min=0, max=img_shape[0] - 1)	388	clamp(min=0, max=img_shape[0] - 1)
389	bboxes = torch.stack([x1, y1, x2, y2], -1)	389	bboxes = torch.stack([x1, y1, x2, y2], -1)
390	det_bboxes.append(bboxes)	390	det_bboxes.append(bboxes)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/free_anchor_retina_head.py+2-2

	@@ -79,7 +79,7 @@ class FreeAnchorRetinaHead(RetinaHead):
79	positive_losses = []	79	positive_losses = []
80	for _, (anchors_, gt_labels_, gt_bboxes_, cls_prob_,	80	for _, (anchors_, gt_labels_, gt_bboxes_, cls_prob_,
81	bbox_preds_) in enumerate(	81	bbox_preds_) in enumerate(
82	- zip(anchors, gt_labels, gt_bboxes, cls_prob, bbox_preds)):	82	+ zip(anchors, gt_labels, gt_bboxes, cls_prob, bbox_preds)):
83	gt_labels_ -= 1	83	gt_labels_ -= 1
84		84
85	with torch.no_grad():	85	with torch.no_grad():
	@@ -197,6 +197,6 @@ class FreeAnchorRetinaHead(RetinaHead):
197		197
198	def negative_bag_loss(self, cls_prob, box_prob):	198	def negative_bag_loss(self, cls_prob, box_prob):
199	prob = cls_prob * (1 - box_prob)	199	prob = cls_prob * (1 - box_prob)
200	- negative_bag_loss = prob*self.gamma F.binary_cross_entropy(	200	+ negative_bag_loss = prob ** self.gamma * F.binary_cross_entropy(
201	prob, torch.zeros_like(prob), reduction='none')	201	prob, torch.zeros_like(prob), reduction='none')
202	return (1 - self.alpha) * negative_bag_loss	202	return (1 - self.alpha) * negative_bag_loss

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/guided_anchor_head.py+29-29

	@@ -108,32 +108,32 @@ class GuidedAnchorHead(AnchorHead):
108	"""	108	"""
109		109
110	def __init__(	110	def __init__(
111	- self,	111	+ self,
112	- num_classes,	112	+ num_classes,
113	- in_channels,	113	+ in_channels,
114	- feat_channels=256,	114	+ feat_channels=256,
115	- octave_base_scale=8,	115	+ octave_base_scale=8,
116	- scales_per_octave=3,	116	+ scales_per_octave=3,
117	- octave_ratios=[0.5, 1.0, 2.0],	117	+ octave_ratios=[0.5, 1.0, 2.0],
118	- anchor_strides=[4, 8, 16, 32, 64],	118	+ anchor_strides=[4, 8, 16, 32, 64],
119	- anchor_base_sizes=None,	119	+ anchor_base_sizes=None,
120	- anchoring_means=(.0, .0, .0, .0),	120	+ anchoring_means=(.0, .0, .0, .0),
121	- anchoring_stds=(1.0, 1.0, 1.0, 1.0),	121	+ anchoring_stds=(1.0, 1.0, 1.0, 1.0),
122	- target_means=(.0, .0, .0, .0),	122	+ target_means=(.0, .0, .0, .0),
123	- target_stds=(1.0, 1.0, 1.0, 1.0),	123	+ target_stds=(1.0, 1.0, 1.0, 1.0),
124	- deformable_groups=4,	124	+ deformable_groups=4,
125	- loc_filter_thr=0.01,	125	+ loc_filter_thr=0.01,
126	- loss_loc=dict(	126	+ loss_loc=dict(
127	- type='FocalLoss',	127	+ type='FocalLoss',
128	- use_sigmoid=True,	128	+ use_sigmoid=True,
129	- gamma=2.0,	129	+ gamma=2.0,
130	- alpha=0.25,	130	+ alpha=0.25,
131	- loss_weight=1.0),	131	+ loss_weight=1.0),
132	- loss_shape=dict(type='BoundedIoULoss', beta=0.2, loss_weight=1.0),	132	+ loss_shape=dict(type='BoundedIoULoss', beta=0.2, loss_weight=1.0),
133	- loss_cls=dict(	133	+ loss_cls=dict(
134	- type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),	134	+ type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
135	- loss_bbox=dict(type='SmoothL1Loss', beta=1.0,	135	+ loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
136	- loss_weight=1.0)): # yapf: disable	136	+ loss_weight=1.0)): # yapf: disable
137	super(AnchorHead, self).__init__()	137	super(AnchorHead, self).__init__()
138	self.in_channels = in_channels	138	self.in_channels = in_channels
139	self.num_classes = num_classes	139	self.num_classes = num_classes
	@@ -141,7 +141,7 @@ class GuidedAnchorHead(AnchorHead):
141	self.octave_base_scale = octave_base_scale	141	self.octave_base_scale = octave_base_scale
142	self.scales_per_octave = scales_per_octave	142	self.scales_per_octave = scales_per_octave
143	self.octave_scales = octave_base_scale * np.array(	143	self.octave_scales = octave_base_scale * np.array(
144	- [2**(i / scales_per_octave) for i in range(scales_per_octave)])	144	+ [2 ** (i / scales_per_octave) for i in range(scales_per_octave)])
145	self.approxs_per_octave = len(self.octave_scales) * len(octave_ratios)	145	self.approxs_per_octave = len(self.octave_scales) * len(octave_ratios)
146	self.octave_ratios = octave_ratios	146	self.octave_ratios = octave_ratios
147	self.anchor_strides = anchor_strides	147	self.anchor_strides = anchor_strides
	@@ -277,7 +277,7 @@ class GuidedAnchorHead(AnchorHead):
277	# inside_flag for a position is true if any anchor in this	277	# inside_flag for a position is true if any anchor in this
278	# position is true	278	# position is true
279	inside_flags = (	279	inside_flags = (
280	- torch.stack(inside_flags_list, 0).sum(dim=0) > 0)	280	+ torch.stack(inside_flags_list, 0).sum(dim=0) > 0)
281	multi_level_flags.append(inside_flags)	281	multi_level_flags.append(inside_flags)
282	inside_flag_list.append(multi_level_flags)	282	inside_flag_list.append(multi_level_flags)
283	return approxs_list, inside_flag_list	283	return approxs_list, inside_flag_list
	@@ -483,7 +483,7 @@ class GuidedAnchorHead(AnchorHead):
483	num_total_pos, num_total_neg) = cls_reg_targets	483	num_total_pos, num_total_neg) = cls_reg_targets
484	num_total_samples = (	484	num_total_samples = (
485	num_total_pos if self.cls_focal_loss else num_total_pos +	485	num_total_pos if self.cls_focal_loss else num_total_pos +
486	- num_total_neg)	486	+ num_total_neg)
487		487
488	# get classification and bbox regression losses	488	# get classification and bbox regression losses
489	losses_cls, losses_bbox = multi_apply(	489	losses_cls, losses_bbox = multi_apply(

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/reppoints_head.py+4-4

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/retina_head.py+1-1

	@@ -59,7 +59,7 @@ class RetinaHead(AnchorHead):
59	self.conv_cfg = conv_cfg	59	self.conv_cfg = conv_cfg
60	self.norm_cfg = norm_cfg	60	self.norm_cfg = norm_cfg
61	octave_scales = np.array(	61	octave_scales = np.array(
62	- [2**(i / scales_per_octave) for i in range(scales_per_octave)])	62	+ [2 ** (i / scales_per_octave) for i in range(scales_per_octave)])
63	anchor_scales = octave_scales * octave_base_scale	63	anchor_scales = octave_scales * octave_base_scale
64	super(RetinaHead, self).__init__(	64	super(RetinaHead, self).__init__(
65	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)	65	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/retina_sepbn_head.py+1-1

	@@ -47,7 +47,7 @@ class RetinaSepBNHead(AnchorHead):
47	self.norm_cfg = norm_cfg	47	self.norm_cfg = norm_cfg
48	self.num_ins = num_ins	48	self.num_ins = num_ins
49	octave_scales = np.array(	49	octave_scales = np.array(
50	- [2**(i / scales_per_octave) for i in range(scales_per_octave)])	50	+ [2 ** (i / scales_per_octave) for i in range(scales_per_octave)])
51	anchor_scales = octave_scales * octave_base_scale	51	anchor_scales = octave_scales * octave_base_scale
52	super(RetinaSepBNHead, self).__init__(	52	super(RetinaSepBNHead, self).__init__(
53	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)	53	num_classes, in_channels, anchor_scales=anchor_scales, **kwargs)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/solo_head.py+28-22

	@@ -25,6 +25,7 @@ from ..utils import bias_init_with_prob, ConvModule
25		25
26	INF = 1e8	26	INF = 1e8
27		27
		28	+
28	def center_of_mass(bitmasks):	29	def center_of_mass(bitmasks):
29	_, h, w = bitmasks.size()	30	_, h, w = bitmasks.size()
30	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)	31	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)
	@@ -37,6 +38,7 @@ def center_of_mass(bitmasks):
37	center_y = m01 / m00	38	center_y = m01 / m00
38	return center_x, center_y	39	return center_x, center_y
39		40
		41	+
40	def points_nms(heat, kernel=2):	42	def points_nms(heat, kernel=2):
41	# kernel must be 2	43	# kernel must be 2
42	hmax = nn.functional.max_pool2d(	44	hmax = nn.functional.max_pool2d(
	@@ -44,6 +46,7 @@ def points_nms(heat, kernel=2):
44	keep = (hmax[:, :, :-1, :-1] == heat).float()	46	keep = (hmax[:, :, :-1, :-1] == heat).float()
45	return heat * keep	47	return heat * keep
46		48
		49	+
47	def dice_loss(input, target):	50	def dice_loss(input, target):
48	input = input.contiguous().view(input.size()[0], -1)	51	input = input.contiguous().view(input.size()[0], -1)
49	target = target.contiguous().view(target.size()[0], -1).float()	52	target = target.contiguous().view(target.size()[0], -1).float()
	@@ -52,7 +55,8 @@ def dice_loss(input, target):
52	b = torch.sum(input * input, 1) + 0.001	55	b = torch.sum(input * input, 1) + 0.001
53	c = torch.sum(target * target, 1) + 0.001	56	c = torch.sum(target * target, 1) + 0.001
54	d = (2 * a) / (b + c)	57	d = (2 * a) / (b + c)
55	- return 1-d	58	+ return 1 - d
		59	+
56		60
57	@HEADS.register_module	61	@HEADS.register_module
58	class SOLOHead(nn.Module):	62	class SOLOHead(nn.Module):
	@@ -123,7 +127,7 @@ class SOLOHead(nn.Module):
123	for seg_num_grid in self.seg_num_grids:	127	for seg_num_grid in self.seg_num_grids:
124	self.solo_ins_list.append(	128	self.solo_ins_list.append(
125	nn.Conv2d(	129	nn.Conv2d(
126	- self.seg_feat_channels, seg_num_grid**2, 1))	130	+ self.seg_feat_channels, seg_num_grid ** 2, 1))
127		131
128	self.solo_cate = nn.Conv2d(	132	self.solo_cate = nn.Conv2d(
129	self.seg_feat_channels, self.cate_out_channels, 3, padding=1)	133	self.seg_feat_channels, self.cate_out_channels, 3, padding=1)
	@@ -134,7 +138,7 @@ class SOLOHead(nn.Module):
134	for m in self.cate_convs:	138	for m in self.cate_convs:
135	normal_init(m.conv, std=0.01)	139	normal_init(m.conv, std=0.01)
136	bias_ins = bias_init_with_prob(0.01)	140	bias_ins = bias_init_with_prob(0.01)
137	- for m in self.solo_ins_list:	141	+ for m in self.solo_ins_list:
138	normal_init(m, std=0.01, bias=bias_ins)	142	normal_init(m, std=0.01, bias=bias_ins)
139	bias_cate = bias_init_with_prob(0.01)	143	bias_cate = bias_init_with_prob(0.01)
140	normal_init(self.solo_cate, std=0.01, bias=bias_cate)	144	normal_init(self.solo_cate, std=0.01, bias=bias_cate)
	@@ -143,16 +147,16 @@ class SOLOHead(nn.Module):
143	new_feats = self.split_feats(feats)	147	new_feats = self.split_feats(feats)
144	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]	148	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]
145	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)	149	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)
146	- ins_pred, cate_pred = multi_apply(self.forward_single, new_feats,	150	+ ins_pred, cate_pred = multi_apply(self.forward_single, new_feats,
147	list(range(len(self.seg_num_grids))),	151	list(range(len(self.seg_num_grids))),
148	eval=eval, upsampled_size=upsampled_size)	152	eval=eval, upsampled_size=upsampled_size)
149	return ins_pred, cate_pred	153	return ins_pred, cate_pred
150		154
151	def split_feats(self, feats):	155	def split_feats(self, feats):
152	- return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),	156	+ return (F.interpolate(feats[0], scale_factor=0.5, mode='bilinear'),
153	- feats[1],	157	+ feats[1],
154	- feats[2],	158	+ feats[2],
155	- feats[3],	159	+ feats[3],
156	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))	160	F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear'))
157		161
158	def forward_single(self, x, idx, eval=False, upsampled_size=None):	162	def forward_single(self, x, idx, eval=False, upsampled_size=None):
	@@ -216,7 +220,6 @@ class SOLOHead(nn.Module):
216	zip(ins_preds_level, ins_ind_labels_level)], 0)	220	zip(ins_preds_level, ins_ind_labels_level)], 0)
217	for ins_preds_level, ins_ind_labels_level in zip(ins_preds, zip(*ins_ind_label_list))]	221	for ins_preds_level, ins_ind_labels_level in zip(ins_preds, zip(*ins_ind_label_list))]
218		222
219	-
220	ins_ind_labels = [	223	ins_ind_labels = [
221	torch.cat([ins_ind_labels_level_img.flatten()	224	torch.cat([ins_ind_labels_level_img.flatten()
222	for ins_ind_labels_level_img in ins_ind_labels_level])	225	for ins_ind_labels_level_img in ins_ind_labels_level])
	@@ -256,10 +259,10 @@ class SOLOHead(nn.Module):
256	loss_cate=loss_cate)	259	loss_cate=loss_cate)
257		260
258	def solo_target_single(self,	261	def solo_target_single(self,
259	- gt_bboxes_raw,	262	+ gt_bboxes_raw,
260	- gt_labels_raw,	263	+ gt_labels_raw,
261	- gt_masks_raw,	264	+ gt_masks_raw,
262	- featmap_sizes=None):	265	+ featmap_sizes=None):
263		266
264	device = gt_labels_raw[0].device	267	device = gt_labels_raw[0].device
265		268
	@@ -296,9 +299,12 @@ class SOLOHead(nn.Module):
296	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0	299	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0
297		300
298	output_stride = stride / 2	301	output_stride = stride / 2
299	- for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels, half_hs, half_ws, center_hs, center_ws, valid_mask_flags):	302	+ for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels,
		303	+ half_hs, half_ws,
		304	+ center_hs, center_ws,
		305	+ valid_mask_flags):
300	if not valid_mask_flag:	306	if not valid_mask_flag:
301	- continue	307	+ continue
302	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)	308	upsampled_size = (featmap_sizes[0][0] * 4, featmap_sizes[0][1] * 4)
303	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))	309	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))
304	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))	310	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))
	@@ -309,17 +315,17 @@ class SOLOHead(nn.Module):
309	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))	315	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))
310	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))	316	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))
311		317
312	- top = max(top_box, coord_h-1)	318	+ top = max(top_box, coord_h - 1)
313	- down = min(down_box, coord_h+1)	319	+ down = min(down_box, coord_h + 1)
314	- left = max(coord_w-1, left_box)	320	+ left = max(coord_w - 1, left_box)
315	- right = min(right_box, coord_w+1)	321	+ right = min(right_box, coord_w + 1)
316		322
317	- cate_label[top:(down+1), left:(right+1)] = gt_label	323	+ cate_label[top:(down + 1), left:(right + 1)] = gt_label
318	# ins	324	# ins
319	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)	325	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)
320	seg_mask = torch.from_numpy(seg_mask).to(device=device)	326	seg_mask = torch.from_numpy(seg_mask).to(device=device)
321	- for i in range(top, down+1):	327	+ for i in range(top, down + 1):
322	- for j in range(left, right+1):	328	+ for j in range(left, right + 1):
323	label = int(i * num_grid + j)	329	label = int(i * num_grid + j)
324	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask	330	ins_label[label, :seg_mask.shape[0], :seg_mask.shape[1]] = seg_mask
325	ins_ind_label[label] = True	331	ins_ind_label[label] = True

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/solov2_head.py+26-16

	@@ -23,8 +23,10 @@ from ..builder import build_loss
23	from ..registry import HEADS	23	from ..registry import HEADS
24	from ..utils import bias_init_with_prob, ConvModule	24	from ..utils import bias_init_with_prob, ConvModule
25	import numpy as np	25	import numpy as np
		26	+
26	INF = 1e8	27	INF = 1e8
27		28
		29	+
28	def center_of_mass(bitmasks):	30	def center_of_mass(bitmasks):
29	_, h, w = bitmasks.size()	31	_, h, w = bitmasks.size()
30	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)	32	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)
	@@ -37,6 +39,7 @@ def center_of_mass(bitmasks):
37	center_y = m01 / m00	39	center_y = m01 / m00
38	return center_x, center_y	40	return center_x, center_y
39		41
		42	+
40	def points_nms(heat, kernel=2):	43	def points_nms(heat, kernel=2):
41	# kernel must be 2	44	# kernel must be 2
42	hmax = nn.functional.max_pool2d(	45	hmax = nn.functional.max_pool2d(
	@@ -44,6 +47,7 @@ def points_nms(heat, kernel=2):
44	keep = (hmax[:, :, :-1, :-1] == heat).float()	47	keep = (hmax[:, :, :-1, :-1] == heat).float()
45	return heat * keep	48	return heat * keep
46		49
		50	+
47	def dice_loss(input, target):	51	def dice_loss(input, target):
48	input = input.contiguous().view(input.size()[0], -1)	52	input = input.contiguous().view(input.size()[0], -1)
49	target = target.contiguous().view(target.size()[0], -1).float()	53	target = target.contiguous().view(target.size()[0], -1).float()
	@@ -52,7 +56,8 @@ def dice_loss(input, target):
52	b = torch.sum(input * input, 1) + 0.001	56	b = torch.sum(input * input, 1) + 0.001
53	c = torch.sum(target * target, 1) + 0.001	57	c = torch.sum(target * target, 1) + 0.001
54	d = (2 * a) / (b + c)	58	d = (2 * a) / (b + c)
55	- return 1-d	59	+ return 1 - d
		60	+
56		61
57	@HEADS.register_module	62	@HEADS.register_module
58	class SOLOv2Head(nn.Module):	63	class SOLOv2Head(nn.Module):
	@@ -150,8 +155,8 @@ class SOLOv2Head(nn.Module):
150	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]	155	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]
151	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)	156	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)
152	cate_pred, kernel_pred = multi_apply(self.forward_single, new_feats,	157	cate_pred, kernel_pred = multi_apply(self.forward_single, new_feats,
153	- list(range(len(self.seg_num_grids))),	158	+ list(range(len(self.seg_num_grids))),
154	- eval=eval, upsampled_size=upsampled_size)	159	+ eval=eval, upsampled_size=upsampled_size)
155	return cate_pred, kernel_pred	160	return cate_pred, kernel_pred
156		161
157	def split_feats(self, feats):	162	def split_feats(self, feats):
	@@ -172,7 +177,7 @@ class SOLOv2Head(nn.Module):
172	x = x.expand([ins_kernel_feat.shape[0], 1, -1, -1])	177	x = x.expand([ins_kernel_feat.shape[0], 1, -1, -1])
173	coord_feat = torch.cat([x, y], 1)	178	coord_feat = torch.cat([x, y], 1)
174	ins_kernel_feat = torch.cat([ins_kernel_feat, coord_feat], 1)	179	ins_kernel_feat = torch.cat([ins_kernel_feat, coord_feat], 1)
175	-	180	+
176	# kernel branch	181	# kernel branch
177	kernel_feat = ins_kernel_feat	182	kernel_feat = ins_kernel_feat
178	seg_num_grid = self.seg_num_grids[idx]	183	seg_num_grid = self.seg_num_grids[idx]
	@@ -195,7 +200,7 @@ class SOLOv2Head(nn.Module):
195	return cate_pred, kernel_pred	200	return cate_pred, kernel_pred
196		201
197	def loss(self,	202	def loss(self,
198	- cate_preds,	203	+ cate_preds,
199	kernel_preds,	204	kernel_preds,
200	ins_pred,	205	ins_pred,
201	gt_bbox_list,	206	gt_bbox_list,
	@@ -203,7 +208,7 @@ class SOLOv2Head(nn.Module):
203	gt_mask_list,	208	gt_mask_list,
204	img_metas,	209	img_metas,
205	cfg,	210	cfg,
206	- gt_bboxes_ignore=None):	211	+ gt_bboxes_ignore=None):
207		212
208	# diff	213	# diff
209	MAX_LEN = 90	214	MAX_LEN = 90
	@@ -258,9 +263,9 @@ class SOLOv2Head(nn.Module):
258	continue	263	continue
259	cur_ins_pred = ins_pred[idx, ...] # this img‘s pred	264	cur_ins_pred = ins_pred[idx, ...] # this img‘s pred
260	H, W = cur_ins_pred.shape[-2:]	265	H, W = cur_ins_pred.shape[-2:]
261	- N, I = kernel_pred.shape	266	+ N, I = kernel_pred.shape
262	cur_ins_pred = cur_ins_pred.unsqueeze(0) # [1, c, msk_pre_h, msk_pre_w] c = N	267	cur_ins_pred = cur_ins_pred.unsqueeze(0) # [1, c, msk_pre_h, msk_pre_w] c = N
263	- kernel_pred = kernel_pred.permute(1, 0).view(I, -1, 1, 1)	268	+ kernel_pred = kernel_pred.permute(1, 0).view(I, -1, 1, 1)
264	cur_ins_pred = F.conv2d(cur_ins_pred, kernel_pred, stride=1).view(-1, H, W) # (n, msk_pre_h, msk_pre_w)	269	cur_ins_pred = F.conv2d(cur_ins_pred, kernel_pred, stride=1).view(-1, H, W) # (n, msk_pre_h, msk_pre_w)
265	b_mask_pred.append(cur_ins_pred)	270	b_mask_pred.append(cur_ins_pred)
266	if len(b_mask_pred) == 0:	271	if len(b_mask_pred) == 0:
	@@ -420,7 +425,7 @@ class SOLOv2Head(nn.Module):
420	seg_pred_list = seg_pred[img_id, ...].unsqueeze(0).float().cpu()	425	seg_pred_list = seg_pred[img_id, ...].unsqueeze(0).float().cpu()
421	kernel_pred_list = [	426	kernel_pred_list = [
422	kernel_preds[i][img_id].permute(1, 2, 0).view(-1, self.kernel_out_channels).detach()	427	kernel_preds[i][img_id].permute(1, 2, 0).view(-1, self.kernel_out_channels).detach()
423	- for i in range(num_levels)	428	+ for i in range(num_levels)
424	]	429	]
425	img_shape = img_metas[img_id]['img_shape']	430	img_shape = img_metas[img_id]['img_shape']
426	scale_factor = img_metas[img_id]['scale_factor']	431	scale_factor = img_metas[img_id]['scale_factor']
	@@ -467,13 +472,18 @@ class SOLOv2Head(nn.Module):
467	n_stage = len(self.seg_num_grids)	472	n_stage = len(self.seg_num_grids)
468	strides[:size_trans[0]] *= self.strides[0]	473	strides[:size_trans[0]] *= self.strides[0]
469	for ind_ in range(1, n_stage):	474	for ind_ in range(1, n_stage):
470	- strides[size_trans[ind_-1]:size_trans[ind_]] *= self.strides[ind_]	475	+ strides[size_trans[ind_ - 1]:size_trans[ind_]] *= self.strides[ind_]
471	strides = strides[inds[:, 0]]	476	strides = strides[inds[:, 0]]
472		477
473	# mask encoding.	478	# mask encoding.
474	I, N = kernel_preds.shape	479	I, N = kernel_preds.shape
475	kernel_preds = kernel_preds.view(I, N, 1, 1)	480	kernel_preds = kernel_preds.view(I, N, 1, 1)
476	- seg_preds = F.conv2d(seg_preds, kernel_preds, stride=1).squeeze(0).sigmoid()	481	+ new_I = ((I // 100) + 1) * 100
		482	+
		483	+ new_kernel_preds = torch.zeros(new_I, N, 1, 1)
		484	+ new_kernel_preds[:I] = kernel_preds
		485	+ seg_preds = F.conv2d(seg_preds.npu().half(), new_kernel_preds.npu().half(), stride=1).squeeze(
		486	+ 0).sigmoid().cpu().float()[:I, :, :]
477	# mask.	487	# mask.
478	seg_masks = seg_preds > cfg.mask_thr	488	seg_masks = seg_preds > cfg.mask_thr
479	sum_masks = seg_masks.sum((1, 2)).float()	489	sum_masks = seg_masks.sum((1, 2)).float()
	@@ -505,7 +515,7 @@ class SOLOv2Head(nn.Module):
505		515
506	# Matrix NMS	516	# Matrix NMS
507	cate_scores = matrix_nms(seg_masks, cate_labels, cate_scores,	517	cate_scores = matrix_nms(seg_masks, cate_labels, cate_scores,
508	- kernel=cfg.kernel,sigma=cfg.sigma, sum_masks=sum_masks)	518	+ kernel=cfg.kernel, sigma=cfg.sigma, sum_masks=sum_masks)
509		519
510	# filter.	520	# filter.
511	keep = cate_scores >= cfg.update_thr	521	keep = cate_scores >= cfg.update_thr
	@@ -524,10 +534,10 @@ class SOLOv2Head(nn.Module):
524	cate_labels = cate_labels[sort_inds]	534	cate_labels = cate_labels[sort_inds]
525		535
526	seg_preds = F.interpolate(seg_preds.unsqueeze(0),	536	seg_preds = F.interpolate(seg_preds.unsqueeze(0),
527	- size=upsampled_size_out,	537	+ size=upsampled_size_out,
528	- mode='bilinear')[:, :, :h, :w]	538	+ mode='bilinear')[:, :, :h, :w]
529	seg_masks = F.interpolate(seg_preds,	539	seg_masks = F.interpolate(seg_preds,
530	- size=ori_shape[:2],	540	+ size=ori_shape[:2],
531	- mode='bilinear').squeeze(0)	541	+ mode='bilinear').squeeze(0)
532	seg_masks = seg_masks > cfg.mask_thr	542	seg_masks = seg_masks > cfg.mask_thr
533	return seg_masks, cate_labels, cate_scores	543	return seg_masks, cate_labels, cate_scores

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/anchor_heads/solov2_light_head.py+32-25

	@@ -25,6 +25,7 @@ from ..utils import bias_init_with_prob, ConvModule
25		25
26	INF = 1e8	26	INF = 1e8
27		27
		28	+
28	def center_of_mass(bitmasks):	29	def center_of_mass(bitmasks):
29	_, h, w = bitmasks.size()	30	_, h, w = bitmasks.size()
30	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)	31	ys = torch.arange(0, h, dtype=torch.float32, device=bitmasks.device)
	@@ -37,6 +38,7 @@ def center_of_mass(bitmasks):
37	center_y = m01 / m00	38	center_y = m01 / m00
38	return center_x, center_y	39	return center_x, center_y
39		40
		41	+
40	def points_nms(heat, kernel=2):	42	def points_nms(heat, kernel=2):
41	# kernel must be 2	43	# kernel must be 2
42	hmax = nn.functional.max_pool2d(	44	hmax = nn.functional.max_pool2d(
	@@ -44,6 +46,7 @@ def points_nms(heat, kernel=2):
44	keep = (hmax[:, :, :-1, :-1] == heat).float()	46	keep = (hmax[:, :, :-1, :-1] == heat).float()
45	return heat * keep	47	return heat * keep
46		48
		49	+
47	def dice_loss(input, target):	50	def dice_loss(input, target):
48	input = input.contiguous().view(input.size()[0], -1)	51	input = input.contiguous().view(input.size()[0], -1)
49	target = target.contiguous().view(target.size()[0], -1).float()	52	target = target.contiguous().view(target.size()[0], -1).float()
	@@ -52,7 +55,8 @@ def dice_loss(input, target):
52	b = torch.sum(input * input, 1) + 0.001	55	b = torch.sum(input * input, 1) + 0.001
53	c = torch.sum(target * target, 1) + 0.001	56	c = torch.sum(target * target, 1) + 0.001
54	d = (2 * a) / (b + c)	57	d = (2 * a) / (b + c)
55	- return 1-d	58	+ return 1 - d
		59	+
56		60
57	@HEADS.register_module	61	@HEADS.register_module
58	class SOLOv2LightHead(nn.Module):	62	class SOLOv2LightHead(nn.Module):
	@@ -150,8 +154,8 @@ class SOLOv2LightHead(nn.Module):
150	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]	154	featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]
151	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)	155	upsampled_size = (featmap_sizes[0][0] * 2, featmap_sizes[0][1] * 2)
152	cate_pred, kernel_pred = multi_apply(self.forward_single, new_feats,	156	cate_pred, kernel_pred = multi_apply(self.forward_single, new_feats,
153	- list(range(len(self.seg_num_grids))),	157	+ list(range(len(self.seg_num_grids))),
154	- eval=eval, upsampled_size=upsampled_size)	158	+ eval=eval, upsampled_size=upsampled_size)
155	return cate_pred, kernel_pred	159	return cate_pred, kernel_pred
156		160
157	def split_feats(self, feats):	161	def split_feats(self, feats):
	@@ -172,7 +176,7 @@ class SOLOv2LightHead(nn.Module):
172	x = x.expand([ins_kernel_feat.shape[0], 1, -1, -1])	176	x = x.expand([ins_kernel_feat.shape[0], 1, -1, -1])
173	coord_feat = torch.cat([x, y], 1)	177	coord_feat = torch.cat([x, y], 1)
174	ins_kernel_feat = torch.cat([ins_kernel_feat, coord_feat], 1)	178	ins_kernel_feat = torch.cat([ins_kernel_feat, coord_feat], 1)
175	-	179	+
176	# kernel branch	180	# kernel branch
177	kernel_feat = ins_kernel_feat	181	kernel_feat = ins_kernel_feat
178	seg_num_grid = self.seg_num_grids[idx]	182	seg_num_grid = self.seg_num_grids[idx]
	@@ -210,7 +214,7 @@ class SOLOv2LightHead(nn.Module):
210	self.solov2_target_single,	214	self.solov2_target_single,
211	gt_bbox_list,	215	gt_bbox_list,
212	gt_label_list,	216	gt_label_list,
213	- gt_mask_list,	217	+ gt_mask_list,
214	mask_feat_size=mask_feat_size)	218	mask_feat_size=mask_feat_size)
215		219
216	# ins	220	# ins
	@@ -283,10 +287,10 @@ class SOLOv2LightHead(nn.Module):
283	loss_cate=loss_cate)	287	loss_cate=loss_cate)
284		288
285	def solov2_target_single(self,	289	def solov2_target_single(self,
286	- gt_bboxes_raw,	290	+ gt_bboxes_raw,
287	- gt_labels_raw,	291	+ gt_labels_raw,
288	- gt_masks_raw,	292	+ gt_masks_raw,
289	- mask_feat_size):	293	+ mask_feat_size):
290		294
291	device = gt_labels_raw[0].device	295	device = gt_labels_raw[0].device
292		296
	@@ -328,9 +332,12 @@ class SOLOv2LightHead(nn.Module):
328	center_ws, center_hs = center_of_mass(gt_masks_pt)	332	center_ws, center_hs = center_of_mass(gt_masks_pt)
329	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0	333	valid_mask_flags = gt_masks_pt.sum(dim=-1).sum(dim=-1) > 0
330	output_stride = 4	334	output_stride = 4
331	- for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels, half_hs, half_ws, center_hs, center_ws, valid_mask_flags):	335	+ for seg_mask, gt_label, half_h, half_w, center_h, center_w, valid_mask_flag in zip(gt_masks, gt_labels,
		336	+ half_hs, half_ws,
		337	+ center_hs, center_ws,
		338	+ valid_mask_flags):
332	if not valid_mask_flag:	339	if not valid_mask_flag:
333	- continue	340	+ continue
334	upsampled_size = (mask_feat_size[0] * 4, mask_feat_size[1] * 4)	341	upsampled_size = (mask_feat_size[0] * 4, mask_feat_size[1] * 4)
335	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))	342	coord_w = int((center_w / upsampled_size[1]) // (1. / num_grid))
336	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))	343	coord_h = int((center_h / upsampled_size[0]) // (1. / num_grid))
	@@ -341,16 +348,16 @@ class SOLOv2LightHead(nn.Module):
341	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))	348	left_box = max(0, int(((center_w - half_w) / upsampled_size[1]) // (1. / num_grid)))
342	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))	349	right_box = min(num_grid - 1, int(((center_w + half_w) / upsampled_size[1]) // (1. / num_grid)))
343		350
344	- top = max(top_box, coord_h-1)	351	+ top = max(top_box, coord_h - 1)
345	- down = min(down_box, coord_h+1)	352	+ down = min(down_box, coord_h + 1)
346	- left = max(coord_w-1, left_box)	353	+ left = max(coord_w - 1, left_box)
347	- right = min(right_box, coord_w+1)	354	+ right = min(right_box, coord_w + 1)
348		355
349	- cate_label[top:(down+1), left:(right+1)] = gt_label	356	+ cate_label[top:(down + 1), left:(right + 1)] = gt_label
350	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)	357	seg_mask = mmcv.imrescale(seg_mask, scale=1. / output_stride)
351	seg_mask = torch.from_numpy(seg_mask).to(device=device)	358	seg_mask = torch.from_numpy(seg_mask).to(device=device)
352	- for i in range(top, down+1):	359	+ for i in range(top, down + 1):
353	- for j in range(left, right+1):	360	+ for j in range(left, right + 1):
354	label = int(i * num_grid + j)	361	label = int(i * num_grid + j)
355		362
356	cur_ins_label = torch.zeros([mask_feat_size[0], mask_feat_size[1]], dtype=torch.uint8,	363	cur_ins_label = torch.zeros([mask_feat_size[0], mask_feat_size[1]], dtype=torch.uint8,
	@@ -381,7 +388,7 @@ class SOLOv2LightHead(nn.Module):
381	seg_pred_list = seg_pred[img_id, ...].unsqueeze(0)	388	seg_pred_list = seg_pred[img_id, ...].unsqueeze(0)
382	kernel_pred_list = [	389	kernel_pred_list = [
383	kernel_preds[i][img_id].permute(1, 2, 0).view(-1, self.kernel_out_channels).detach()	390	kernel_preds[i][img_id].permute(1, 2, 0).view(-1, self.kernel_out_channels).detach()
384	- for i in range(num_levels)	391	+ for i in range(num_levels)
385	]	392	]
386	img_shape = img_metas[img_id]['img_shape']	393	img_shape = img_metas[img_id]['img_shape']
387	scale_factor = img_metas[img_id]['scale_factor']	394	scale_factor = img_metas[img_id]['scale_factor']
	@@ -430,7 +437,7 @@ class SOLOv2LightHead(nn.Module):
430	n_stage = len(self.seg_num_grids)	437	n_stage = len(self.seg_num_grids)
431	strides[:size_trans[0]] *= self.strides[0]	438	strides[:size_trans[0]] *= self.strides[0]
432	for ind_ in range(1, n_stage):	439	for ind_ in range(1, n_stage):
433	- strides[size_trans[ind_-1]:size_trans[ind_]] *= self.strides[ind_]	440	+ strides[size_trans[ind_ - 1]:size_trans[ind_]] *= self.strides[ind_]
434	strides = strides[inds[:, 0]]	441	strides = strides[inds[:, 0]]
435		442
436	# mask encoding.	443	# mask encoding.
	@@ -468,7 +475,7 @@ class SOLOv2LightHead(nn.Module):
468		475
469	# Matrix NMS	476	# Matrix NMS
470	cate_scores = matrix_nms(seg_masks, cate_labels, cate_scores,	477	cate_scores = matrix_nms(seg_masks, cate_labels, cate_scores,
471	- kernel=cfg.kernel,sigma=cfg.sigma, sum_masks=sum_masks)	478	+ kernel=cfg.kernel, sigma=cfg.sigma, sum_masks=sum_masks)
472		479
473	# filter.	480	# filter.
474	keep = cate_scores >= cfg.update_thr	481	keep = cate_scores >= cfg.update_thr
	@@ -487,10 +494,10 @@ class SOLOv2LightHead(nn.Module):
487	cate_labels = cate_labels[sort_inds]	494	cate_labels = cate_labels[sort_inds]
488		495
489	seg_preds = F.interpolate(seg_preds.unsqueeze(0),	496	seg_preds = F.interpolate(seg_preds.unsqueeze(0),
490	- size=upsampled_size_out,	497	+ size=upsampled_size_out,
491	- mode='bilinear')[:, :, :h, :w]	498	+ mode='bilinear')[:, :, :h, :w]
492	seg_masks = F.interpolate(seg_preds,	499	seg_masks = F.interpolate(seg_preds,
493	- size=ori_shape[:2],	500	+ size=ori_shape[:2],
494	- mode='bilinear').squeeze(0)	501	+ mode='bilinear').squeeze(0)
495	seg_masks = seg_masks > cfg.mask_thr	502	seg_masks = seg_masks > cfg.mask_thr
496	return seg_masks, cate_labels, cate_scores	503	return seg_masks, cate_labels, cate_scores

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/backbones/hrnet.py+1-1

	@@ -148,7 +148,7 @@ class HRModule(nn.Module):
148	bias=False),	148	bias=False),
149	build_norm_layer(self.norm_cfg, in_channels[i])[1],	149	build_norm_layer(self.norm_cfg, in_channels[i])[1],
150	nn.Upsample(	150	nn.Upsample(
151	- scale_factor=2**(j - i), mode='nearest')))	151	+ scale_factor=2 ** (j - i), mode='nearest')))
152	elif j == i:	152	elif j == i:
153	fuse_layer.append(None)	153	fuse_layer.append(None)
154	else:	154	else:

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/backbones/resnet.py+3-3

	@@ -424,7 +424,7 @@ class ResNet(nn.Module):
424	dilation = dilations[i]	424	dilation = dilations[i]
425	dcn = self.dcn if self.stage_with_dcn[i] else None	425	dcn = self.dcn if self.stage_with_dcn[i] else None
426	gcb = self.gcb if self.stage_with_gcb[i] else None	426	gcb = self.gcb if self.stage_with_gcb[i] else None
427	- planes = 64 * 2**i	427	+ planes = 64 * 2 ** i
428	res_layer = make_res_layer(	428	res_layer = make_res_layer(
429	self.block,	429	self.block,
430	self.inplanes,	430	self.inplanes,
	@@ -447,8 +447,8 @@ class ResNet(nn.Module):
447		447
448	self._freeze_stages()	448	self._freeze_stages()
449		449
450	- self.feat_dim = self.block.expansion * 64 * 2**(	450	+ self.feat_dim = self.block.expansion * 64 * 2 ** (
451	- len(self.stage_blocks) - 1)	451	+ len(self.stage_blocks) - 1)
452		452
453	@property	453	@property
454	def norm1(self):	454	def norm1(self):

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/backbones/resnext.py+1-1

	@@ -212,7 +212,7 @@ class ResNeXt(ResNet):
212	dilation = self.dilations[i]	212	dilation = self.dilations[i]
213	dcn = self.dcn if self.stage_with_dcn[i] else None	213	dcn = self.dcn if self.stage_with_dcn[i] else None
214	gcb = self.gcb if self.stage_with_gcb[i] else None	214	gcb = self.gcb if self.stage_with_gcb[i] else None
215	- planes = 64 * 2**i	215	+ planes = 64 * 2 ** i
216	res_layer = make_res_layer(	216	res_layer = make_res_layer(
217	self.block,	217	self.block,
218	self.inplanes,	218	self.inplanes,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/bbox_heads/bbox_head.py+2-2

	@@ -186,7 +186,7 @@ class BBoxHead(nn.Module):
186		186
187	return det_bboxes, det_labels	187	return det_bboxes, det_labels
188		188
189	- @force_fp32(apply_to=('bbox_preds', ))	189	+ @force_fp32(apply_to=('bbox_preds',))
190	def refine_bboxes(self, rois, labels, bbox_preds, pos_is_gts, img_metas):	190	def refine_bboxes(self, rois, labels, bbox_preds, pos_is_gts, img_metas):
191	"""Refine bboxes during training.	191	"""Refine bboxes during training.
192		192
	@@ -264,7 +264,7 @@ class BBoxHead(nn.Module):
264		264
265	return bboxes_list	265	return bboxes_list
266		266
267	- @force_fp32(apply_to=('bbox_pred', ))	267	+ @force_fp32(apply_to=('bbox_pred',))
268	def regress_by_class(self, rois, label, bbox_pred, img_meta):	268	def regress_by_class(self, rois, label, bbox_pred, img_meta):
269	"""Regress the bbox for the predicted class. Used in Cascade R-CNN.	269	"""Regress the bbox for the predicted class. Used in Cascade R-CNN.
270		270

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/bbox_heads/convfc_bbox_head.py+2-2

	@@ -91,7 +91,7 @@ class ConvFCBBoxHead(BBoxHead):
91	self.fc_cls = nn.Linear(self.cls_last_dim, self.num_classes)	91	self.fc_cls = nn.Linear(self.cls_last_dim, self.num_classes)
92	if self.with_reg:	92	if self.with_reg:
93	out_dim_reg = (4 if self.reg_class_agnostic else 4 *	93	out_dim_reg = (4 if self.reg_class_agnostic else 4 *
94	- self.num_classes)	94	+ self.num_classes)
95	self.fc_reg = nn.Linear(self.reg_last_dim, out_dim_reg)	95	self.fc_reg = nn.Linear(self.reg_last_dim, out_dim_reg)
96		96
97	def _add_conv_fc_branch(self,	97	def _add_conv_fc_branch(self,
	@@ -125,7 +125,7 @@ class ConvFCBBoxHead(BBoxHead):
125	# for shared branch, only consider self.with_avg_pool	125	# for shared branch, only consider self.with_avg_pool
126	# for separated branches, also consider self.num_shared_fcs	126	# for separated branches, also consider self.num_shared_fcs
127	if (is_shared	127	if (is_shared
128	- or self.num_shared_fcs == 0) and not self.with_avg_pool:	128	+ or self.num_shared_fcs == 0) and not self.with_avg_pool:
129	last_layer_dim *= self.roi_feat_area	129	last_layer_dim *= self.roi_feat_area
130	for i in range(num_branch_fcs):	130	for i in range(num_branch_fcs):
131	fc_in_channels = (	131	fc_in_channels = (

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/base.py+2-2

	@@ -37,7 +37,7 @@ class BaseDetector(nn.Module, metaclass=ABCMeta):
37	@property	37	@property
38	def with_mask_feat_head(self):	38	def with_mask_feat_head(self):
39	return hasattr(self, 'mask_feat_head') and \	39	return hasattr(self, 'mask_feat_head') and \
40	- self.mask_feat_head is not None	40	+ self.mask_feat_head is not None
41		41
42	@property	42	@property
43	def with_shared_head(self):	43	def with_shared_head(self):
	@@ -147,7 +147,7 @@ class BaseDetector(nn.Module, metaclass=ABCMeta):
147	else:	147	else:
148	return self.aug_test(imgs, img_metas, **kwargs)	148	return self.aug_test(imgs, img_metas, **kwargs)
149		149
150	- @auto_fp16(apply_to=('img', ))	150	+ @auto_fp16(apply_to=('img',))
151	def forward(self, img, img_meta, return_loss=True, **kwargs):	151	def forward(self, img, img_meta, return_loss=True, **kwargs):
152	"""	152	"""
153	Calls either forward_train or forward_test depending on whether	153	Calls either forward_train or forward_test depending on whether

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/cascade_rcnn.py+2-2

	@@ -139,7 +139,7 @@ class CascadeRCNN(BaseDetector, RPNTestMixin):
139	# rpn	139	# rpn
140	if self.with_rpn:	140	if self.with_rpn:
141	rpn_outs = self.rpn_head(x)	141	rpn_outs = self.rpn_head(x)
142	- outs = outs + (rpn_outs, )	142	+ outs = outs + (rpn_outs,)
143	proposals = torch.randn(1000, 4).cuda()	143	proposals = torch.randn(1000, 4).cuda()
144	# bbox heads	144	# bbox heads
145	rois = bbox2roi([proposals])	145	rois = bbox2roi([proposals])
	@@ -160,7 +160,7 @@ class CascadeRCNN(BaseDetector, RPNTestMixin):
160	if self.with_shared_head:	160	if self.with_shared_head:
161	mask_feats = self.shared_head(mask_feats)	161	mask_feats = self.shared_head(mask_feats)
162	mask_pred = self.mask_head[i](mask_feats)	162	mask_pred = self.mask_head[i](mask_feats)
163	- outs = outs + (mask_pred, )	163	+ outs = outs + (mask_pred,)
164	return outs	164	return outs
165		165
166	def forward_train(self,	166	def forward_train(self,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/double_head_rcnn.py+1-1

	@@ -33,7 +33,7 @@ class DoubleHeadRCNN(TwoStageDetector):
33	# rpn	33	# rpn
34	if self.with_rpn:	34	if self.with_rpn:
35	rpn_outs = self.rpn_head(x)	35	rpn_outs = self.rpn_head(x)
36	- outs = outs + (rpn_outs, )	36	+ outs = outs + (rpn_outs,)
37	proposals = torch.randn(1000, 4).cuda()	37	proposals = torch.randn(1000, 4).cuda()
38	# bbox head	38	# bbox head
39	rois = bbox2roi([proposals])	39	rois = bbox2roi([proposals])

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/grid_rcnn.py+1-1

	@@ -101,7 +101,7 @@ class GridRCNN(TwoStageDetector):
101	# rpn	101	# rpn
102	if self.with_rpn:	102	if self.with_rpn:
103	rpn_outs = self.rpn_head(x)	103	rpn_outs = self.rpn_head(x)
104	- outs = outs + (rpn_outs, )	104	+ outs = outs + (rpn_outs,)
105	proposals = torch.randn(1000, 4).cuda()	105	proposals = torch.randn(1000, 4).cuda()
106	# bbox head	106	# bbox head
107	rois = bbox2roi([proposals])	107	rois = bbox2roi([proposals])

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/htc.py+3-3

	@@ -175,7 +175,7 @@ class HybridTaskCascade(CascadeRCNN):
175	# rpn	175	# rpn
176	if self.with_rpn:	176	if self.with_rpn:
177	rpn_outs = self.rpn_head(x)	177	rpn_outs = self.rpn_head(x)
178	- outs = outs + (rpn_outs, )	178	+ outs = outs + (rpn_outs,)
179	proposals = torch.randn(1000, 4).cuda()	179	proposals = torch.randn(1000, 4).cuda()
180	# semantic head	180	# semantic head
181	if self.with_semantic:	181	if self.with_semantic:
	@@ -205,7 +205,7 @@ class HybridTaskCascade(CascadeRCNN):
205	mask_pred, last_feat = mask_head(mask_feats, last_feat)	205	mask_pred, last_feat = mask_head(mask_feats, last_feat)
206	else:	206	else:
207	mask_pred = mask_head(mask_feats)	207	mask_pred = mask_head(mask_feats)
208	- outs = outs + (mask_pred, )	208	+ outs = outs + (mask_pred,)
209	return outs	209	return outs
210		210
211	def forward_train(self,	211	def forward_train(self,
	@@ -499,7 +499,7 @@ class HybridTaskCascade(CascadeRCNN):
499	mask_semantic_feat = self.semantic_roi_extractor(	499	mask_semantic_feat = self.semantic_roi_extractor(
500	[semantic_feat], mask_rois)	500	[semantic_feat], mask_rois)
501	if mask_semantic_feat.shape[-2:] != mask_feats.shape[	501	if mask_semantic_feat.shape[-2:] != mask_feats.shape[
502	- -2:]:	502	+ -2:]:
503	mask_semantic_feat = F.adaptive_avg_pool2d(	503	mask_semantic_feat = F.adaptive_avg_pool2d(
504	mask_semantic_feat, mask_feats.shape[-2:])	504	mask_semantic_feat, mask_feats.shape[-2:])
505	mask_feats += mask_semantic_feat	505	mask_feats += mask_semantic_feat

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/single_stage_ins.py+3-4

	@@ -91,7 +91,7 @@ class SingleStageInsDetector(BaseDetector):
91	if self.with_mask_feat_head:	91	if self.with_mask_feat_head:
92	mask_feat_pred = self.mask_feat_head(	92	mask_feat_pred = self.mask_feat_head(
93	x[self.mask_feat_head.	93	x[self.mask_feat_head.
94	- start_level:self.mask_feat_head.end_level + 1])	94	+ start_level:self.mask_feat_head.end_level + 1])
95	loss_inputs = outs + (mask_feat_pred, gt_bboxes, gt_labels, gt_masks, img_metas, self.train_cfg)	95	loss_inputs = outs + (mask_feat_pred, gt_bboxes, gt_labels, gt_masks, img_metas, self.train_cfg)
96	else:	96	else:
97	loss_inputs = outs + (gt_bboxes, gt_labels, gt_masks, img_metas, self.train_cfg)	97	loss_inputs = outs + (gt_bboxes, gt_labels, gt_masks, img_metas, self.train_cfg)
	@@ -102,19 +102,18 @@ class SingleStageInsDetector(BaseDetector):
102	def simple_test(self, img, img_meta, rescale=False):	102	def simple_test(self, img, img_meta, rescale=False):
103	# diff	103	# diff
104	img = img.npu()	104	img = img.npu()
105	-
106	x = self.extract_feat(img)	105	x = self.extract_feat(img)
107	outs = self.bbox_head(x, eval=True)	106	outs = self.bbox_head(x, eval=True)
108	if self.with_mask_feat_head:	107	if self.with_mask_feat_head:
109	mask_feat_pred = self.mask_feat_head(	108	mask_feat_pred = self.mask_feat_head(
110	x[self.mask_feat_head.	109	x[self.mask_feat_head.
111	- start_level:self.mask_feat_head.end_level + 1])	110	+ start_level:self.mask_feat_head.end_level + 1])
112	seg_inputs = outs + (mask_feat_pred, img_meta, self.test_cfg, rescale)	111	seg_inputs = outs + (mask_feat_pred, img_meta, self.test_cfg, rescale)
113	else:	112	else:
114	seg_inputs = outs + (img_meta, self.test_cfg, rescale)	113	seg_inputs = outs + (img_meta, self.test_cfg, rescale)
115		114
116	seg_result = self.bbox_head.get_seg(*seg_inputs)	115	seg_result = self.bbox_head.get_seg(*seg_inputs)
117	- return seg_result	116	+ return seg_result
118		117
119	def aug_test(self, imgs, img_metas, rescale=False):	118	def aug_test(self, imgs, img_metas, rescale=False):
120	raise NotImplementedError	119	raise NotImplementedError

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/solov2.py+1-1

	@@ -28,4 +28,4 @@ class SOLOv2(SingleStageInsDetector):
28	test_cfg=None,	28	test_cfg=None,
29	pretrained=None):	29	pretrained=None):
30	super(SOLOv2, self).__init__(backbone, neck, bbox_head, mask_feat_head, train_cfg,	30	super(SOLOv2, self).__init__(backbone, neck, bbox_head, mask_feat_head, train_cfg,
31	- test_cfg, pretrained)	31	+ test_cfg, pretrained)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/test_mixins.py+0-4

	@@ -27,9 +27,7 @@ if sys.version_info >= (3, 7):
27		27
28		28
29	class RPNTestMixin(object):	29	class RPNTestMixin(object):
30	-
31	if sys.version_info >= (3, 7):	30	if sys.version_info >= (3, 7):
32	-
33	async def async_test_rpn(self, x, img_meta, rpn_test_cfg):	31	async def async_test_rpn(self, x, img_meta, rpn_test_cfg):
34	sleep_interval = rpn_test_cfg.pop("async_sleep_interval", 0.025)	32	sleep_interval = rpn_test_cfg.pop("async_sleep_interval", 0.025)
35	async with completed(	33	async with completed(
	@@ -72,7 +70,6 @@ class RPNTestMixin(object):
72		70
73		71
74	class BBoxTestMixin(object):	72	class BBoxTestMixin(object):
75	-
76	if sys.version_info >= (3, 7):	73	if sys.version_info >= (3, 7):
77		74
78	async def async_test_bboxes(self,	75	async def async_test_bboxes(self,
	@@ -172,7 +169,6 @@ class BBoxTestMixin(object):
172		169
173		170
174	class MaskTestMixin(object):	171	class MaskTestMixin(object):
175	-
176	if sys.version_info >= (3, 7):	172	if sys.version_info >= (3, 7):
177		173
178	async def async_test_mask(self,	174	async def async_test_mask(self,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/detectors/two_stage.py+2-2

	@@ -119,7 +119,7 @@ class TwoStageDetector(BaseDetector, RPNTestMixin, BBoxTestMixin,
119	# rpn	119	# rpn
120	if self.with_rpn:	120	if self.with_rpn:
121	rpn_outs = self.rpn_head(x)	121	rpn_outs = self.rpn_head(x)
122	- outs = outs + (rpn_outs, )	122	+ outs = outs + (rpn_outs,)
123	proposals = torch.randn(1000, 4).cuda()	123	proposals = torch.randn(1000, 4).cuda()
124	# bbox head	124	# bbox head
125	rois = bbox2roi([proposals])	125	rois = bbox2roi([proposals])
	@@ -138,7 +138,7 @@ class TwoStageDetector(BaseDetector, RPNTestMixin, BBoxTestMixin,
138	if self.with_shared_head:	138	if self.with_shared_head:
139	mask_feats = self.shared_head(mask_feats)	139	mask_feats = self.shared_head(mask_feats)
140	mask_pred = self.mask_head(mask_feats)	140	mask_pred = self.mask_head(mask_feats)
141	- outs = outs + (mask_pred, )	141	+ outs = outs + (mask_pred,)
142	return outs	142	return outs
143		143
144	def forward_train(self,	144	def forward_train(self,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/losses/accuracy.py+2-2

	@@ -18,7 +18,7 @@ import torch.nn as nn
18	def accuracy(pred, target, topk=1):	18	def accuracy(pred, target, topk=1):
19	assert isinstance(topk, (int, tuple))	19	assert isinstance(topk, (int, tuple))
20	if isinstance(topk, int):	20	if isinstance(topk, int):
21	- topk = (topk, )	21	+ topk = (topk,)
22	return_single = True	22	return_single = True
23	else:	23	else:
24	return_single = False	24	return_single = False
	@@ -37,7 +37,7 @@ def accuracy(pred, target, topk=1):
37		37
38	class Accuracy(nn.Module):	38	class Accuracy(nn.Module):
39		39
40	- def __init__(self, topk=(1, )):	40	+ def __init__(self, topk=(1,)):
41	super().__init__()	41	super().__init__()
42	self.topk = topk	42	self.topk = topk
43		43

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/losses/balanced_l1_loss.py+1-1

	@@ -31,7 +31,7 @@ def balanced_l1_loss(pred,
31	assert pred.size() == target.size() and target.numel() > 0	31	assert pred.size() == target.size() and target.numel() > 0
32		32
33	diff = torch.abs(pred - target)	33	diff = torch.abs(pred - target)
34	- b = np.e**(gamma / alpha) - 1	34	+ b = np.e ** (gamma / alpha) - 1
35	loss = torch.where(	35	loss = torch.where(
36	diff < beta, alpha / b *	36	diff < beta, alpha / b *
37	(b * diff + 1) * torch.log(b * diff / beta + 1) - alpha * diff,	37	(b * diff + 1) * torch.log(b * diff / beta + 1) - alpha * diff,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/losses/focal_loss.py+2-1

	@@ -15,7 +15,7 @@
15	import torch	15	import torch
16	import torch.nn as nn	16	import torch.nn as nn
17	import torch.nn.functional as F	17	import torch.nn.functional as F
18	-#from mmcv.ops import sigmoid_focal_loss as _sigmoid_focal_loss	18	+# from mmcv.ops import sigmoid_focal_loss as _sigmoid_focal_loss
19		19
20	from ..builder import LOSSES	20	from ..builder import LOSSES
21	from .utils import weight_reduce_loss	21	from .utils import weight_reduce_loss
	@@ -55,6 +55,7 @@ def py_sigmoid_focal_loss(pred,
55	loss = weight_reduce_loss(loss, weight, reduction, avg_factor)	55	loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
56	return loss	56	return loss
57		57
		58	+
58	def sigmoid_focal_loss(pred,	59	def sigmoid_focal_loss(pred,
59	target,	60	target,
60	weight=None,	61	weight=None,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/losses/ghm_loss.py+2-2

	@@ -94,7 +94,7 @@ class GHMC(nn.Module):
94	if num_in_bin > 0:	94	if num_in_bin > 0:
95	if mmt > 0:	95	if mmt > 0:
96	self.acc_sum[i] = mmt * self.acc_sum[i] \	96	self.acc_sum[i] = mmt * self.acc_sum[i] \
97	- + (1 - mmt) * num_in_bin	97	+ + (1 - mmt) * num_in_bin
98	weights[inds] = tot / self.acc_sum[i]	98	weights[inds] = tot / self.acc_sum[i]
99	else:	99	else:
100	weights[inds] = tot / num_in_bin	100	weights[inds] = tot / num_in_bin
	@@ -173,7 +173,7 @@ class GHMR(nn.Module):
173	n += 1	173	n += 1
174	if mmt > 0:	174	if mmt > 0:
175	self.acc_sum[i] = mmt * self.acc_sum[i] \	175	self.acc_sum[i] = mmt * self.acc_sum[i] \
176	- + (1 - mmt) * num_in_bin	176	+ + (1 - mmt) * num_in_bin
177	weights[inds] = tot / self.acc_sum[i]	177	weights[inds] = tot / self.acc_sum[i]
178	else:	178	else:
179	weights[inds] = tot / num_in_bin	179	weights[inds] = tot / num_in_bin

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/mask_heads/fcn_mask_head.py+1-1

	@@ -125,7 +125,7 @@ class FCNMaskHead(nn.Module):
125	gt_masks, rcnn_train_cfg)	125	gt_masks, rcnn_train_cfg)
126	return mask_targets	126	return mask_targets
127		127
128	- @force_fp32(apply_to=('mask_pred', ))	128	+ @force_fp32(apply_to=('mask_pred',))
129	def loss(self, mask_pred, mask_targets, labels):	129	def loss(self, mask_pred, mask_targets, labels):
130	loss = dict()	130	loss = dict()
131	if self.class_agnostic:	131	if self.class_agnostic:

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/mask_heads/fused_semantic_head.py+1-1

	@@ -112,7 +112,7 @@ class FusedSemanticHead(nn.Module):
112	x = self.conv_embedding(x)	112	x = self.conv_embedding(x)
113	return mask_pred, x	113	return mask_pred, x
114		114
115	- @force_fp32(apply_to=('mask_pred', ))	115	+ @force_fp32(apply_to=('mask_pred',))
116	def loss(self, mask_pred, labels):	116	def loss(self, mask_pred, labels):
117	labels = labels.squeeze(1).long()	117	labels = labels.squeeze(1).long()
118	loss_semantic_seg = self.criterion(mask_pred, labels)	118	loss_semantic_seg = self.criterion(mask_pred, labels)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/mask_heads/grid_head.py+8-8

	@@ -267,7 +267,7 @@ class GridHead(nn.Module):
267	1 - y_idx / (self.grid_size - 1)))	267	1 - y_idx / (self.grid_size - 1)))
268		268
269	radius = rcnn_train_cfg.pos_radius	269	radius = rcnn_train_cfg.pos_radius
270	- radius2 = radius**2	270	+ radius2 = radius ** 2
271	for i in range(num_rois):	271	for i in range(num_rois):
272	# ignore small bboxes	272	# ignore small bboxes
273	if (pos_bbox_ws[i] <= self.grid_size	273	if (pos_bbox_ws[i] <= self.grid_size
	@@ -277,9 +277,9 @@ class GridHead(nn.Module):
277	for j in range(self.grid_points):	277	for j in range(self.grid_points):
278	factor_x, factor_y = factors[j]	278	factor_x, factor_y = factors[j]
279	gridpoint_x = factor_x * pos_gt_bboxes[i, 0] + (	279	gridpoint_x = factor_x * pos_gt_bboxes[i, 0] + (
280	- 1 - factor_x) * pos_gt_bboxes[i, 2]	280	+ 1 - factor_x) * pos_gt_bboxes[i, 2]
281	gridpoint_y = factor_y * pos_gt_bboxes[i, 1] + (	281	gridpoint_y = factor_y * pos_gt_bboxes[i, 1] + (
282	- 1 - factor_y) * pos_gt_bboxes[i, 3]	282	+ 1 - factor_y) * pos_gt_bboxes[i, 3]
283		283
284	cx = int((gridpoint_x - pos_bboxes[i, 0]) / pos_bbox_ws[i] *	284	cx = int((gridpoint_x - pos_bboxes[i, 0]) / pos_bbox_ws[i] *
285	map_size)	285	map_size)
	@@ -289,7 +289,7 @@ class GridHead(nn.Module):
289	for x in range(cx - radius, cx + radius + 1):	289	for x in range(cx - radius, cx + radius + 1):
290	for y in range(cy - radius, cy + radius + 1):	290	for y in range(cy - radius, cy + radius + 1):
291	if x >= 0 and x < map_size and y >= 0 and y < map_size:	291	if x >= 0 and x < map_size and y >= 0 and y < map_size:
292	- if (x - cx)2 + (y - cy)2 <= radius2:	292	+ if (x - cx) 2 + (y - cy) 2 <= radius2:
293	targets[i, j, y, x] = 1	293	targets[i, j, y, x] = 1
294	# reduce the target heatmap size by a half	294	# reduce the target heatmap size by a half
295	# proposed in Grid R-CNN Plus (https://arxiv.org/abs/1906.05688).	295	# proposed in Grid R-CNN Plus (https://arxiv.org/abs/1906.05688).
	@@ -356,16 +356,16 @@ class GridHead(nn.Module):
356	# voting of all grid points on some boundary	356	# voting of all grid points on some boundary
357	bboxes_x1 = (abs_xs[:, x1_inds] * pred_scores[:, x1_inds]).sum(	357	bboxes_x1 = (abs_xs[:, x1_inds] * pred_scores[:, x1_inds]).sum(
358	dim=1, keepdim=True) / (	358	dim=1, keepdim=True) / (
359	- pred_scores[:, x1_inds].sum(dim=1, keepdim=True))	359	+ pred_scores[:, x1_inds].sum(dim=1, keepdim=True))
360	bboxes_y1 = (abs_ys[:, y1_inds] * pred_scores[:, y1_inds]).sum(	360	bboxes_y1 = (abs_ys[:, y1_inds] * pred_scores[:, y1_inds]).sum(
361	dim=1, keepdim=True) / (	361	dim=1, keepdim=True) / (
362	- pred_scores[:, y1_inds].sum(dim=1, keepdim=True))	362	+ pred_scores[:, y1_inds].sum(dim=1, keepdim=True))
363	bboxes_x2 = (abs_xs[:, x2_inds] * pred_scores[:, x2_inds]).sum(	363	bboxes_x2 = (abs_xs[:, x2_inds] * pred_scores[:, x2_inds]).sum(
364	dim=1, keepdim=True) / (	364	dim=1, keepdim=True) / (
365	- pred_scores[:, x2_inds].sum(dim=1, keepdim=True))	365	+ pred_scores[:, x2_inds].sum(dim=1, keepdim=True))
366	bboxes_y2 = (abs_ys[:, y2_inds] * pred_scores[:, y2_inds]).sum(	366	bboxes_y2 = (abs_ys[:, y2_inds] * pred_scores[:, y2_inds]).sum(
367	dim=1, keepdim=True) / (	367	dim=1, keepdim=True) / (
368	- pred_scores[:, y2_inds].sum(dim=1, keepdim=True))	368	+ pred_scores[:, y2_inds].sum(dim=1, keepdim=True))
369		369
370	bbox_res = torch.cat(	370	bbox_res = torch.cat(
371	[bboxes_x1, bboxes_y1, bboxes_x2, bboxes_y2, cls_scores], dim=1)	371	[bboxes_x1, bboxes_y1, bboxes_x2, bboxes_y2, cls_scores], dim=1)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/mask_heads/mask_feat_head.py+2-2

	@@ -63,7 +63,7 @@ class MaskFeatHead(nn.Module):
63		63
64	for j in range(i):	64	for j in range(i):
65	if j == 0:	65	if j == 0:
66	- chn = self.in_channels+2 if i==3 else self.in_channels	66	+ chn = self.in_channels + 2 if i == 3 else self.in_channels
67	one_conv = ConvModule(	67	one_conv = ConvModule(
68	chn,	68	chn,
69	self.out_channels,	69	self.out_channels,
	@@ -126,7 +126,7 @@ class MaskFeatHead(nn.Module):
126	x = x.expand([input_feat.shape[0], 1, -1, -1])	126	x = x.expand([input_feat.shape[0], 1, -1, -1])
127	coord_feat = torch.cat([x, y], 1)	127	coord_feat = torch.cat([x, y], 1)
128	input_p = torch.cat([input_p, coord_feat], 1)	128	input_p = torch.cat([input_p, coord_feat], 1)
129	-	129	+
130	feature_add_all_level += self.convs_all_levels[i](input_p)	130	feature_add_all_level += self.convs_all_levels[i](input_p)
131		131
132	feature_pred = self.conv_pred(feature_add_all_level)	132	feature_pred = self.conv_pred(feature_add_all_level)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/mask_heads/maskiou_head.py+6-6

	@@ -102,7 +102,7 @@ class MaskIoUHead(nn.Module):
102	mask_iou = self.fc_mask_iou(x)	102	mask_iou = self.fc_mask_iou(x)
103	return mask_iou	103	return mask_iou
104		104
105	- @force_fp32(apply_to=('mask_iou_pred', ))	105	+ @force_fp32(apply_to=('mask_iou_pred',))
106	def loss(self, mask_iou_pred, mask_iou_targets):	106	def loss(self, mask_iou_pred, mask_iou_targets):
107	pos_inds = mask_iou_targets > 0	107	pos_inds = mask_iou_targets > 0
108	if pos_inds.sum() > 0:	108	if pos_inds.sum() > 0:
	@@ -112,7 +112,7 @@ class MaskIoUHead(nn.Module):
112	loss_mask_iou = mask_iou_pred * 0	112	loss_mask_iou = mask_iou_pred * 0
113	return dict(loss_mask_iou=loss_mask_iou)	113	return dict(loss_mask_iou=loss_mask_iou)
114		114
115	- @force_fp32(apply_to=('mask_pred', ))	115	+ @force_fp32(apply_to=('mask_pred',))
116	def get_target(self, sampling_results, gt_masks, mask_pred, mask_targets,	116	def get_target(self, sampling_results, gt_masks, mask_pred, mask_targets,
117	rcnn_train_cfg):	117	rcnn_train_cfg):
118	"""Compute target of mask IoU.	118	"""Compute target of mask IoU.
	@@ -159,7 +159,7 @@ class MaskIoUHead(nn.Module):
159	gt_full_areas = mask_targets.sum((-1, -2)) / (area_ratios + 1e-7)	159	gt_full_areas = mask_targets.sum((-1, -2)) / (area_ratios + 1e-7)
160		160
161	mask_iou_targets = overlap_areas / (	161	mask_iou_targets = overlap_areas / (
162	- mask_pred_areas + gt_full_areas - overlap_areas)	162	+ mask_pred_areas + gt_full_areas - overlap_areas)
163	return mask_iou_targets	163	return mask_iou_targets
164		164
165	def _get_area_ratio(self, pos_proposals, pos_assigned_gt_inds, gt_masks):	165	def _get_area_ratio(self, pos_proposals, pos_assigned_gt_inds, gt_masks):
	@@ -180,15 +180,15 @@ class MaskIoUHead(nn.Module):
180	gt_mask_in_proposal = gt_mask[y1:y2 + 1, x1:x2 + 1]	180	gt_mask_in_proposal = gt_mask[y1:y2 + 1, x1:x2 + 1]
181		181
182	ratio = gt_mask_in_proposal.sum() / (	182	ratio = gt_mask_in_proposal.sum() / (
183	- gt_instance_mask_area[pos_assigned_gt_inds[i]] + 1e-7)	183	+ gt_instance_mask_area[pos_assigned_gt_inds[i]] + 1e-7)
184	area_ratios.append(ratio)	184	area_ratios.append(ratio)
185	area_ratios = torch.from_numpy(np.stack(area_ratios)).float().to(	185	area_ratios = torch.from_numpy(np.stack(area_ratios)).float().to(
186	pos_proposals.device)	186	pos_proposals.device)
187	else:	187	else:
188	- area_ratios = pos_proposals.new_zeros((0, ))	188	+ area_ratios = pos_proposals.new_zeros((0,))
189	return area_ratios	189	return area_ratios
190		190
191	- @force_fp32(apply_to=('mask_iou_pred', ))	191	+ @force_fp32(apply_to=('mask_iou_pred',))
192	def get_mask_scores(self, mask_iou_pred, det_bboxes, det_labels):	192	def get_mask_scores(self, mask_iou_pred, det_bboxes, det_labels):
193	"""Get the mask scores.	193	"""Get the mask scores.
194		194

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/necks/hrfpn.py+2-2

	@@ -94,7 +94,7 @@ class HRFPN(nn.Module):
94	outs = [inputs[0]]	94	outs = [inputs[0]]
95	for i in range(1, self.num_ins):	95	for i in range(1, self.num_ins):
96	outs.append(	96	outs.append(
97	- F.interpolate(inputs[i], scale_factor=2**i, mode='bilinear'))	97	+ F.interpolate(inputs[i], scale_factor=2 ** i, mode='bilinear'))
98	out = torch.cat(outs, dim=1)	98	out = torch.cat(outs, dim=1)
99	if out.requires_grad and self.with_cp:	99	if out.requires_grad and self.with_cp:
100	out = checkpoint(self.reduction_conv, out)	100	out = checkpoint(self.reduction_conv, out)
	@@ -102,7 +102,7 @@ class HRFPN(nn.Module):
102	out = self.reduction_conv(out)	102	out = self.reduction_conv(out)
103	outs = [out]	103	outs = [out]
104	for i in range(1, self.num_outs):	104	for i in range(1, self.num_outs):
105	- outs.append(self.pooling(out, kernel_size=2i, stride=2i))	105	+ outs.append(self.pooling(out, kernel_size=2 i, stride=2 i))
106	outputs = []	106	outputs = []
107		107
108	for i in range(self.num_outs):	108	for i in range(self.num_outs):

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/plugins/generalized_attention.py+45-45

	@@ -135,15 +135,15 @@ class GeneralizedAttention(nn.Module):
135	for iy in range(max_len):	135	for iy in range(max_len):
136	for ix in range(max_len):	136	for ix in range(max_len):
137	local_constraint_map[	137	local_constraint_map[
138	- iy, ix,	138	+ iy, ix,
139	- max((iy - self.spatial_range) //	139	+ max((iy - self.spatial_range) //
140	- self.kv_stride, 0):min((iy + self.spatial_range +	140	+ self.kv_stride, 0):min((iy + self.spatial_range +
141	- 1) // self.kv_stride +	141	+ 1) // self.kv_stride +
142	- 1, max_len),	142	+ 1, max_len),
143	- max((ix - self.spatial_range) //	143	+ max((ix - self.spatial_range) //
144	- self.kv_stride, 0):min((ix + self.spatial_range +	144	+ self.kv_stride, 0):min((ix + self.spatial_range +
145	- 1) // self.kv_stride +	145	+ 1) // self.kv_stride +
146	- 1, max_len)] = 0	146	+ 1, max_len)] = 0
147		147
148	self.local_constraint_map = nn.Parameter(	148	self.local_constraint_map = nn.Parameter(
149	torch.from_numpy(local_constraint_map).byte(),	149	torch.from_numpy(local_constraint_map).byte(),
	@@ -196,7 +196,7 @@ class GeneralizedAttention(nn.Module):
196	feat_range = torch.arange(0, feat_dim / 4).cuda(device)	196	feat_range = torch.arange(0, feat_dim / 4).cuda(device)
197		197
198	dim_mat = torch.Tensor([wave_length]).cuda(device)	198	dim_mat = torch.Tensor([wave_length]).cuda(device)
199	- dim_mat = dim_mat*((4. / feat_dim) feat_range)	199	+ dim_mat = dim_mat ** ((4. / feat_dim) * feat_range)
200	dim_mat = dim_mat.view((1, 1, -1))	200	dim_mat = dim_mat.view((1, 1, -1))
201		201
202	embedding_x = torch.cat(	202	embedding_x = torch.cat(
	@@ -237,15 +237,15 @@ class GeneralizedAttention(nn.Module):
237	h, w, h_kv, w_kv, self.q_stride, self.kv_stride,	237	h, w, h_kv, w_kv, self.q_stride, self.kv_stride,
238	x_input.device, self.position_embedding_dim)	238	x_input.device, self.position_embedding_dim)
239	# (n, num_heads, w, w_kv, dim)	239	# (n, num_heads, w, w_kv, dim)
240	- position_feat_x = self.appr_geom_fc_x(position_embed_x).\	240	+ position_feat_x = self.appr_geom_fc_x(position_embed_x). \
241	- view(1, w, w_kv, num_heads, self.qk_embed_dim).\	241	+ view(1, w, w_kv, num_heads, self.qk_embed_dim). \
242	- permute(0, 3, 1, 2, 4).\	242	+ permute(0, 3, 1, 2, 4). \
243	repeat(n, 1, 1, 1, 1)	243	repeat(n, 1, 1, 1, 1)
244		244
245	# (n, num_heads, h, h_kv, dim)	245	# (n, num_heads, h, h_kv, dim)
246	- position_feat_y = self.appr_geom_fc_y(position_embed_y).\	246	+ position_feat_y = self.appr_geom_fc_y(position_embed_y). \
247	- view(1, h, h_kv, num_heads, self.qk_embed_dim).\	247	+ view(1, h, h_kv, num_heads, self.qk_embed_dim). \
248	- permute(0, 3, 1, 2, 4).\	248	+ permute(0, 3, 1, 2, 4). \
249	repeat(n, 1, 1, 1, 1)	249	repeat(n, 1, 1, 1, 1)
250		250
251	position_feat_x /= math.sqrt(2)	251	position_feat_x /= math.sqrt(2)
	@@ -253,11 +253,11 @@ class GeneralizedAttention(nn.Module):
253		253
254	# accelerate for saliency only	254	# accelerate for saliency only
255	if (np.sum(self.attention_type) == 1) and self.attention_type[2]:	255	if (np.sum(self.attention_type) == 1) and self.attention_type[2]:
256	- appr_bias = self.appr_bias.\	256	+ appr_bias = self.appr_bias. \
257	- view(1, num_heads, 1, self.qk_embed_dim).\	257	+ view(1, num_heads, 1, self.qk_embed_dim). \
258	repeat(n, 1, 1, 1)	258	repeat(n, 1, 1, 1)
259		259
260	- energy = torch.matmul(appr_bias, proj_key).\	260	+ energy = torch.matmul(appr_bias, proj_key). \
261	view(n, num_heads, 1, h_kv * w_kv)	261	view(n, num_heads, 1, h_kv * w_kv)
262		262
263	h = 1	263	h = 1
	@@ -281,35 +281,35 @@ class GeneralizedAttention(nn.Module):
281	# attention_type[3]: bias - position	281	# attention_type[3]: bias - position
282	if self.attention_type[0] or self.attention_type[2]:	282	if self.attention_type[0] or self.attention_type[2]:
283	if self.attention_type[0] and self.attention_type[2]:	283	if self.attention_type[0] and self.attention_type[2]:
284	- appr_bias = self.appr_bias.\	284	+ appr_bias = self.appr_bias. \
285	view(1, num_heads, 1, self.qk_embed_dim)	285	view(1, num_heads, 1, self.qk_embed_dim)
286	- energy = torch.matmul(proj_query + appr_bias, proj_key).\	286	+ energy = torch.matmul(proj_query + appr_bias, proj_key). \
287	view(n, num_heads, h, w, h_kv, w_kv)	287	view(n, num_heads, h, w, h_kv, w_kv)
288		288
289	elif self.attention_type[0]:	289	elif self.attention_type[0]:
290	- energy = torch.matmul(proj_query, proj_key).\	290	+ energy = torch.matmul(proj_query, proj_key). \
291	view(n, num_heads, h, w, h_kv, w_kv)	291	view(n, num_heads, h, w, h_kv, w_kv)
292		292
293	elif self.attention_type[2]:	293	elif self.attention_type[2]:
294	- appr_bias = self.appr_bias.\	294	+ appr_bias = self.appr_bias. \
295	- view(1, num_heads, 1, self.qk_embed_dim).\	295	+ view(1, num_heads, 1, self.qk_embed_dim). \
296	repeat(n, 1, 1, 1)	296	repeat(n, 1, 1, 1)
297		297
298	- energy += torch.matmul(appr_bias, proj_key).\	298	+ energy += torch.matmul(appr_bias, proj_key). \
299	view(n, num_heads, 1, 1, h_kv, w_kv)	299	view(n, num_heads, 1, 1, h_kv, w_kv)
300		300
301	if self.attention_type[1] or self.attention_type[3]:	301	if self.attention_type[1] or self.attention_type[3]:
302	if self.attention_type[1] and self.attention_type[3]:	302	if self.attention_type[1] and self.attention_type[3]:
303	- geom_bias = self.geom_bias.\	303	+ geom_bias = self.geom_bias. \
304	view(1, num_heads, 1, self.qk_embed_dim)	304	view(1, num_heads, 1, self.qk_embed_dim)
305		305
306	- proj_query_reshape = (proj_query + geom_bias).\	306	+ proj_query_reshape = (proj_query + geom_bias). \
307	view(n, num_heads, h, w, self.qk_embed_dim)	307	view(n, num_heads, h, w, self.qk_embed_dim)
308		308
309	energy_x = torch.matmul(	309	energy_x = torch.matmul(
310	proj_query_reshape.permute(0, 1, 3, 2, 4),	310	proj_query_reshape.permute(0, 1, 3, 2, 4),
311	position_feat_x.permute(0, 1, 2, 4, 3))	311	position_feat_x.permute(0, 1, 2, 4, 3))
312	- energy_x = energy_x.\	312	+ energy_x = energy_x. \
313	permute(0, 1, 3, 2, 4).unsqueeze(4)	313	permute(0, 1, 3, 2, 4).unsqueeze(4)
314		314
315	energy_y = torch.matmul(	315	energy_y = torch.matmul(
	@@ -320,13 +320,13 @@ class GeneralizedAttention(nn.Module):
320	energy += energy_x + energy_y	320	energy += energy_x + energy_y
321		321
322	elif self.attention_type[1]:	322	elif self.attention_type[1]:
323	- proj_query_reshape = proj_query.\	323	+ proj_query_reshape = proj_query. \
324	view(n, num_heads, h, w, self.qk_embed_dim)	324	view(n, num_heads, h, w, self.qk_embed_dim)
325	- proj_query_reshape = proj_query_reshape.\	325	+ proj_query_reshape = proj_query_reshape. \
326	permute(0, 1, 3, 2, 4)	326	permute(0, 1, 3, 2, 4)
327	- position_feat_x_reshape = position_feat_x.\	327	+ position_feat_x_reshape = position_feat_x. \
328	permute(0, 1, 2, 4, 3)	328	permute(0, 1, 2, 4, 3)
329	- position_feat_y_reshape = position_feat_y.\	329	+ position_feat_y_reshape = position_feat_y. \
330	permute(0, 1, 2, 4, 3)	330	permute(0, 1, 2, 4, 3)
331		331
332	energy_x = torch.matmul(proj_query_reshape,	332	energy_x = torch.matmul(proj_query_reshape,
	@@ -340,14 +340,14 @@ class GeneralizedAttention(nn.Module):
340	energy += energy_x + energy_y	340	energy += energy_x + energy_y
341		341
342	elif self.attention_type[3]:	342	elif self.attention_type[3]:
343	- geom_bias = self.geom_bias.\	343	+ geom_bias = self.geom_bias. \
344	- view(1, num_heads, self.qk_embed_dim, 1).\	344	+ view(1, num_heads, self.qk_embed_dim, 1). \
345	repeat(n, 1, 1, 1)	345	repeat(n, 1, 1, 1)
346		346
347	- position_feat_x_reshape = position_feat_x.\	347	+ position_feat_x_reshape = position_feat_x. \
348	- view(n, num_heads, w*w_kv, self.qk_embed_dim)	348	+ view(n, num_heads, w * w_kv, self.qk_embed_dim)
349		349
350	- position_feat_y_reshape = position_feat_y.\	350	+ position_feat_y_reshape = position_feat_y. \
351	view(n, num_heads, h * h_kv, self.qk_embed_dim)	351	view(n, num_heads, h * h_kv, self.qk_embed_dim)
352		352
353	energy_x = torch.matmul(position_feat_x_reshape, geom_bias)	353	energy_x = torch.matmul(position_feat_x_reshape, geom_bias)
	@@ -362,9 +362,9 @@ class GeneralizedAttention(nn.Module):
362		362
363	if self.spatial_range >= 0:	363	if self.spatial_range >= 0:
364	cur_local_constraint_map = \	364	cur_local_constraint_map = \
365	- self.local_constraint_map[:h, :w, :h_kv, :w_kv].\	365	+ self.local_constraint_map[:h, :w, :h_kv, :w_kv]. \
366	- contiguous().\	366	+ contiguous(). \
367	- view(1, 1, hw, h_kvw_kv)	367	+ view(1, 1, h * w, h_kv * w_kv)
368		368
369	energy = energy.masked_fill_(cur_local_constraint_map,	369	energy = energy.masked_fill_(cur_local_constraint_map,
370	float('-inf'))	370	float('-inf'))
	@@ -372,13 +372,13 @@ class GeneralizedAttention(nn.Module):
372	attention = F.softmax(energy, 3)	372	attention = F.softmax(energy, 3)
373		373
374	proj_value = self.value_conv(x_kv)	374	proj_value = self.value_conv(x_kv)
375	- proj_value_reshape = proj_value.\	375	+ proj_value_reshape = proj_value. \
376	- view((n, num_heads, self.v_dim, h_kv * w_kv)).\	376	+ view((n, num_heads, self.v_dim, h_kv * w_kv)). \
377	permute(0, 1, 3, 2)	377	permute(0, 1, 3, 2)
378		378
379	- out = torch.matmul(attention, proj_value_reshape).\	379	+ out = torch.matmul(attention, proj_value_reshape). \
380	- permute(0, 1, 3, 2).\	380	+ permute(0, 1, 3, 2). \
381	- contiguous().\	381	+ contiguous(). \
382	view(n, self.v_dim * self.num_heads, h, w)	382	view(n, self.v_dim * self.num_heads, h, w)
383		383
384	out = self.proj_conv(out)	384	out = self.proj_conv(out)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/plugins/non_local.py+1-1

	@@ -90,7 +90,7 @@ class NonLocal2D(nn.Module):
90	pairwise_weight = torch.matmul(theta_x, phi_x)	90	pairwise_weight = torch.matmul(theta_x, phi_x)
91	if self.use_scale:	91	if self.use_scale:
92	# theta_x.shape[-1] is `self.inter_channels`	92	# theta_x.shape[-1] is `self.inter_channels`
93	- pairwise_weight /= theta_x.shape[-1]**0.5	93	+ pairwise_weight /= theta_x.shape[-1] ** 0.5
94	pairwise_weight = pairwise_weight.softmax(dim=-1)	94	pairwise_weight = pairwise_weight.softmax(dim=-1)
95	return pairwise_weight	95	return pairwise_weight
96		96

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/roi_extractors/single_level.py+1-1

	@@ -100,7 +100,7 @@ class SingleRoIExtractor(nn.Module):
100	new_rois = torch.stack((rois[:, 0], x1, y1, x2, y2), dim=-1)	100	new_rois = torch.stack((rois[:, 0], x1, y1, x2, y2), dim=-1)
101	return new_rois	101	return new_rois
102		102
103	- @force_fp32(apply_to=('feats', ), out_fp16=True)	103	+ @force_fp32(apply_to=('feats',), out_fp16=True)
104	def forward(self, feats, rois, roi_scale_factor=None):	104	def forward(self, feats, rois, roi_scale_factor=None):
105	if len(feats) == 1:	105	if len(feats) == 1:
106	return self.roi_layers[0](feats[0], rois)	106	return self.roi_layers[0](feats[0], rois)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/models/shared_heads/res_layer.py+2-2

	@@ -42,8 +42,8 @@ class ResLayer(nn.Module):
42	self.fp16_enabled = False	42	self.fp16_enabled = False
43	block, stage_blocks = ResNet.arch_settings[depth]	43	block, stage_blocks = ResNet.arch_settings[depth]
44	stage_block = stage_blocks[stage]	44	stage_block = stage_blocks[stage]
45	- planes = 64 * 2**stage	45	+ planes = 64 * 2 ** stage
46	- inplanes = 64 * 2*(stage - 1) block.expansion	46	+ inplanes = 64 * 2 ** (stage - 1) * block.expansion
47		47
48	res_layer = make_res_layer(	48	res_layer = make_res_layer(
49	block,	49	block,

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/ops/context_block.py+1-1

	@@ -30,7 +30,7 @@ class ContextBlock(nn.Module):
30	inplanes,	30	inplanes,
31	ratio,	31	ratio,
32	pooling_type='att',	32	pooling_type='att',
33	- fusion_types=('channel_add', )):	33	+ fusion_types=('channel_add',)):
34	super(ContextBlock, self).__init__()	34	super(ContextBlock, self).__init__()
35	assert pooling_type in ['avg', 'att']	35	assert pooling_type in ['avg', 'att']
36	assert isinstance(fusion_types, (list, tuple))	36	assert isinstance(fusion_types, (list, tuple))

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/ops/dcn/deform_conv.py+3-1

	@@ -18,6 +18,7 @@ from torch.autograd import Function
18	from torch.nn.modules.utils import _pair, _single	18	from torch.nn.modules.utils import _pair, _single
19	import math	19	import math
20		20
		21	+
21	class ModulatedDeformConv2dFunction(Function):	22	class ModulatedDeformConv2dFunction(Function):
22		23
23	@staticmethod	24	@staticmethod
	@@ -198,10 +199,11 @@ class ModulatedDeformConvPack(ModulatedDeformConv2d):
198	self.stride, self.padding, self.dilation,	199	self.stride, self.padding, self.dilation,
199	self.groups, self.deformable_groups)	200	self.groups, self.deformable_groups)
200		201
		202	+
201	DCNv2 = ModulatedDeformConvPack	203	DCNv2 = ModulatedDeformConvPack
202		204
203	if __name__ == "__main__":	205	if __name__ == "__main__":
204	- x = torch.randn(2,32,4,4)	206	+ x = torch.randn(2, 32, 4, 4)
205	model = DCNv2(32, 32, 1)	207	model = DCNv2(32, 32, 1)
206		208
207	torch.npu.set_device(0)	209	torch.npu.set_device(0)

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/ops/masked_conv/masked_conv.py+1-1

	@@ -69,7 +69,7 @@ class MaskedConv2dFunction(Function):
69	@staticmethod	69	@staticmethod
70	@once_differentiable	70	@once_differentiable
71	def backward(ctx, grad_output):	71	def backward(ctx, grad_output):
72	- return (None, ) * 5	72	+ return (None,) * 5
73		73
74		74
75	masked_conv2d = MaskedConv2dFunction.apply	75	masked_conv2d = MaskedConv2dFunction.apply

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/ops/nms/nms_wrapper.py+1-0

	@@ -15,6 +15,7 @@
15	import numpy as np	15	import numpy as np
16	import torch	16	import torch
17		17
		18	+
18	# from . import nms_cpu, nms_cuda	19	# from . import nms_cpu, nms_cuda
19	# from .soft_nms_cpu import soft_nms_cpu	20	# from .soft_nms_cpu import soft_nms_cpu
20		21

MPyTorch/contrib/cv/detection/SOLOv2/mmdet/utils/flops_counter.py+14-16

	@@ -82,21 +82,21 @@ def get_model_complexity_info(model,
82		82
83	def flops_to_string(flops, units='GMac', precision=2):	83	def flops_to_string(flops, units='GMac', precision=2):
84	if units is None:	84	if units is None:
85	- if flops // 10**9 > 0:	85	+ if flops // 10 ** 9 > 0:
86	- return str(round(flops / 10.**9, precision)) + ' GMac'	86	+ return str(round(flops / 10. ** 9, precision)) + ' GMac'
87	- elif flops // 10**6 > 0:	87	+ elif flops // 10 ** 6 > 0:
88	- return str(round(flops / 10.**6, precision)) + ' MMac'	88	+ return str(round(flops / 10. ** 6, precision)) + ' MMac'
89	- elif flops // 10**3 > 0:	89	+ elif flops // 10 ** 3 > 0:
90	- return str(round(flops / 10.**3, precision)) + ' KMac'	90	+ return str(round(flops / 10. ** 3, precision)) + ' KMac'
91	else:	91	else:
92	return str(flops) + ' Mac'	92	return str(flops) + ' Mac'
93	else:	93	else:
94	if units == 'GMac':	94	if units == 'GMac':
95	- return str(round(flops / 10.**9, precision)) + ' ' + units	95	+ return str(round(flops / 10. ** 9, precision)) + ' ' + units
96	elif units == 'MMac':	96	elif units == 'MMac':
97	- return str(round(flops / 10.**6, precision)) + ' ' + units	97	+ return str(round(flops / 10. ** 6, precision)) + ' ' + units
98	elif units == 'KMac':	98	elif units == 'KMac':
99	- return str(round(flops / 10.**3, precision)) + ' ' + units	99	+ return str(round(flops / 10. ** 3, precision)) + ' ' + units
100	else:	100	else:
101	return str(flops) + ' Mac'	101	return str(flops) + ' Mac'
102		102
	@@ -114,10 +114,10 @@ def params_to_string(params_num):
114	>>> params_to_string(3e-9)	114	>>> params_to_string(3e-9)
115	'3e-09'	115	'3e-09'
116	"""	116	"""
117	- if params_num // 10**6 > 0:	117	+ if params_num // 10 ** 6 > 0:
118	- return str(round(params_num / 10**6, 2)) + ' M'	118	+ return str(round(params_num / 10 ** 6, 2)) + ' M'
119	- elif params_num // 10**3:	119	+ elif params_num // 10 ** 3:
120	- return str(round(params_num / 10**3, 2)) + ' k'	120	+ return str(round(params_num / 10 ** 3, 2)) + ' k'
121	else:	121	else:
122	return str(params_num)	122	return str(params_num)
123		123
	@@ -237,7 +237,6 @@ def reset_flops_count(self):
237		237
238		238
239	def add_flops_mask(module, mask):	239	def add_flops_mask(module, mask):
240	-
241	def add_flops_mask_func(module):	240	def add_flops_mask_func(module):
242	if isinstance(module, torch.nn.Conv2d):	241	if isinstance(module, torch.nn.Conv2d):
243	module.__mask__ = mask	242	module.__mask__ = mask
	@@ -319,7 +318,7 @@ def deconv_flops_counter_hook(conv_module, input, output):
319		318
320	filters_per_channel = out_channels // groups	319	filters_per_channel = out_channels // groups
321	conv_per_position_flops = (	320	conv_per_position_flops = (
322	- kernel_height * kernel_width * in_channels * filters_per_channel)	321	+ kernel_height * kernel_width * in_channels * filters_per_channel)
323		322
324	active_elements_count = batch_size * input_height * input_width	323	active_elements_count = batch_size * input_height * input_width
325	overall_conv_flops = conv_per_position_flops * active_elements_count	324	overall_conv_flops = conv_per_position_flops * active_elements_count
	@@ -362,7 +361,6 @@ def conv_flops_counter_hook(conv_module, input, output):
362	bias_flops = 0	361	bias_flops = 0
363		362
364	if conv_module.bias is not None:	363	if conv_module.bias is not None:
365	-
366	bias_flops = out_channels * active_elements_count	364	bias_flops = out_channels * active_elements_count
367		365
368	overall_flops = overall_conv_flops + bias_flops	366	overall_flops = overall_conv_flops + bias_flops

MPyTorch/contrib/cv/detection/SOLOv2/paddlepaddle/__init__.py+1-1

	@@ -15,4 +15,4 @@
15	# Author: Acer Zhang	15	# Author: Acer Zhang
16	# Datetime: 2021/9/1	16	# Datetime: 2021/9/1
17	# Copyright belongs to the author.	17	# Copyright belongs to the author.
18	-# Please indicate the source for reprinting.	18	+# Please indicate the source for reprinting.

MPyTorch/contrib/cv/detection/SOLOv2/setup.py+1-3

	@@ -12,7 +12,7 @@
12	# See the License for the specific language governing permissions and	12	# See the License for the specific language governing permissions and
13	# limitations under the License.	13	# limitations under the License.
14		14
15	-#!/usr/bin/env python	15	+# !/usr/bin/env python
16	# -- coding: utf-8 --	16	# -- coding: utf-8 --
17	import os	17	import os
18	import platform	18	import platform
	@@ -48,7 +48,6 @@ version_file = 'mmdet/version.py'
48		48
49		49
50	def get_git_hash():	50	def get_git_hash():
51	-
52	def _minimal_ext_cmd(cmd):	51	def _minimal_ext_cmd(cmd):
53	# construct minimal environment	52	# construct minimal environment
54	env = {}	53	env = {}
	@@ -131,7 +130,6 @@ def make_cuda_ext(name, module, sources):
131	define_macros=define_macros,	130	define_macros=define_macros,
132	extra_compile_args=extra_compile_args)	131	extra_compile_args=extra_compile_args)
133		132
134	-
135	# define_macros = []	133	# define_macros = []
136	# extra_compile_args = {'cxx': []}	134	# extra_compile_args = {'cxx': []}
137	#	135	#

MPyTorch/contrib/cv/detection/SOLOv2/test/train_eval_1p.sh+2-2

	@@ -11,7 +11,7 @@ batch_size=1
11	# 训练使用的npu卡数	11	# 训练使用的npu卡数
12	export RANK_SIZE=1	12	export RANK_SIZE=1
13	data_path=""	13	data_path=""
14	-MODEL="./work_dirs/solov2_release_r50_fpn_8gpu_1x/latest.pth"	14	+MODEL="./work_dirs/solov2_release_r50_fpn_8gpu_1x/epoch_12.pth"
15	device_id=0	15	device_id=0
16		16
17	#参数校验，不需要修改	17	#参数校验，不需要修改
	@@ -70,7 +70,7 @@ end_time=$(date +%s)
70	e2e_time=$(( $end_time - $start_time ))	70	e2e_time=$(( $end_time - $start_time ))
71		71
72	# 输出训练精度,需要模型审视修改 # eval.log \| awk -F ',' '{print $1}' \| awk '{print $2}' \| awk ' END {print}'	72	# 输出训练精度,需要模型审视修改 # eval.log \| awk -F ',' '{print $1}' \| awk '{print $2}' \| awk ' END {print}'
73	-train_accuracy=`grep -a 'maxDets' $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log\|awk -F " " '{print $13}'\|head -n 1`	73	+train_accuracy=`grep -a 'maxDets' $cur_path/output/${ASCEND_DEVICE_ID}/eval_${ASCEND_DEVICE_ID}.log\|awk -F " " '{print $13}'\|head -n 1`
74	# 打印，不需要修改	74	# 打印，不需要修改
75	echo "Final Train Accuracy : ${train_accuracy}"	75	echo "Final Train Accuracy : ${train_accuracy}"
76	echo "E2E Training Duration sec : $e2e_time"	76	echo "E2E Training Duration sec : $e2e_time"

MPyTorch/contrib/cv/detection/SOLOv2/test/train_full_1p.sh+1-1

	@@ -14,7 +14,7 @@ Network="SOLOv2"
14		14
15	#训练batch_size,,需要模型审视修改	15	#训练batch_size,,需要模型审视修改
16	batch_size=2	16	batch_size=2
17	-device_id=1	17	+device_id=0
18		18
19	#参数校验，不需要修改	19	#参数校验，不需要修改
20	for para in $*	20	for para in $*

MPyTorch/contrib/cv/detection/SOLOv2/test/train_full_8p.sh+1-1

	@@ -14,7 +14,7 @@ Network="SOLOv2"
14		14
15	#训练batch_size,,需要模型审视修改	15	#训练batch_size,,需要模型审视修改
16	batch_size=16	16	batch_size=16
17	-device_id=1	17	+device_id=0
18		18
19	#参数校验，不需要修改	19	#参数校验，不需要修改
20	for para in $*	20	for para in $*

MPyTorch/contrib/cv/detection/SOLOv2/test/train_performance_1p.sh+2-5

	@@ -14,7 +14,7 @@ Network="SOLOv2"
14		14
15	#训练batch_size,,需要模型审视修改	15	#训练batch_size,,需要模型审视修改
16	batch_size=2	16	batch_size=2
17	-device_id=1	17	+device_id=0
18		18
19	#参数校验，不需要修改	19	#参数校验，不需要修改
20	for para in $*	20	for para in $*
	@@ -68,10 +68,7 @@ fi
68	export NPUID=0	68	export NPUID=0
69	export RANK=0	69	export RANK=0
70	python3.7 tools/train.py configs/solov2/solov2_r50_fpn_8gpu_1x.py --opt-level $apex --autoscale-lr --seed 0 --total_epochs 1 \	70	python3.7 tools/train.py configs/solov2/solov2_r50_fpn_8gpu_1x.py --opt-level $apex --autoscale-lr --seed 0 --total_epochs 1 \
71	- --data_root=$data_path --gpu-ids 0 > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	71	+ --data_root=$data_path --gpu-ids 0 --train_performance=True > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
72	-wait
73	-python3.7 tools/test_ins.py configs/solov2/solov2_r50_fpn_8gpu_1x.py work_dirs/solov2_release_r50_fpn_8gpu_1x/latest.pth --show \
74	- --out results_solo.pkl --eval segm --data_root=$data_path >> ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
75	wait	72	wait
76		73
77	#训练结束时间，不需要修改	74	#训练结束时间，不需要修改

MPyTorch/contrib/cv/detection/SOLOv2/test/train_performance_8p.sh+3-6

	@@ -14,7 +14,7 @@ Network="SOLOv2"
14		14
15	#训练batch_size,,需要模型审视修改	15	#训练batch_size,,需要模型审视修改
16	batch_size=16	16	batch_size=16
17	-device_id=1	17	+device_id=0
18		18
19	#参数校验，不需要修改	19	#参数校验，不需要修改
20	for para in $*	20	for para in $*
	@@ -82,7 +82,7 @@ do
82	--gpus 8 \	82	--gpus 8 \
83	--autoscale-lr \	83	--autoscale-lr \
84	--seed 0 \	84	--seed 0 \
85	- --data_root=$data_path \	85	+ --data_root=$data_path --train_performance=True \
86	--total_epochs 1 > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	86	--total_epochs 1 > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
87	else	87	else
88	python3.7 ./tools/train.py configs/solov2/solov2_r50_fpn_8gpu_1x.py \	88	python3.7 ./tools/train.py configs/solov2/solov2_r50_fpn_8gpu_1x.py \
	@@ -91,14 +91,11 @@ do
91	--gpus 8 \	91	--gpus 8 \
92	--autoscale-lr \	92	--autoscale-lr \
93	--seed 0 \	93	--seed 0 \
94	- --data_root=$data_path \	94	+ --data_root=$data_path --train_performance=True \
95	--total_epochs 1 > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &	95	--total_epochs 1 > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
96	fi	96	fi
97	done	97	done
98	wait	98	wait
99	-python3.7 tools/test_ins.py configs/solov2/solov2_r50_fpn_8gpu_1x.py work_dirs/solov2_release_r50_fpn_8gpu_1x/latest.pth --show \
100	- --out results_solo.pkl --eval segm --data_root=$data_path >> ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
101	-wait
102	#训练结束时间，不需要修改	99	#训练结束时间，不需要修改
103	end_time=$(date +%s)	100	end_time=$(date +%s)
104	e2e_time=$(( $end_time - $start_time ))	101	e2e_time=$(( $end_time - $start_time ))

MPyTorch/contrib/cv/detection/SOLOv2/tests/test_assigner.py+1-1

	@@ -118,7 +118,7 @@ def test_max_iou_assigner_with_empty_boxes():
118	# Test with gt_labels	118	# Test with gt_labels
119	assign_result = self.assign(bboxes, gt_bboxes, gt_labels=gt_labels)	119	assign_result = self.assign(bboxes, gt_bboxes, gt_labels=gt_labels)
120	assert len(assign_result.gt_inds) == 0	120	assert len(assign_result.gt_inds) == 0
121	- assert tuple(assign_result.labels.shape) == (0, )	121	+ assert tuple(assign_result.labels.shape) == (0,)
122		122
123	# Test without gt_labels	123	# Test without gt_labels
124	assign_result = self.assign(bboxes, gt_bboxes, gt_labels=None)	124	assign_result = self.assign(bboxes, gt_bboxes, gt_labels=None)

MPyTorch/contrib/cv/detection/SOLOv2/tests/test_async.py+0-1

	@@ -73,7 +73,6 @@ class MaskRCNNDetector:
73		73
74		74
75	class AsyncInferenceTestCase(AsyncTestCase):	75	class AsyncInferenceTestCase(AsyncTestCase):
76	-
77	if sys.version_info >= (3, 7):	76	if sys.version_info >= (3, 7):
78		77
79	async def test_simple_inference(self):	78	async def test_simple_inference(self):

MPyTorch/contrib/cv/detection/SOLOv2/tests/test_heads.py+5-5

	@@ -54,7 +54,7 @@ def test_anchor_head_loss():
54		54
55	# Anchor head expects a multiple levels of features per image	55	# Anchor head expects a multiple levels of features per image
56	feat = [	56	feat = [
57	- torch.rand(1, 1, s // (2(i + 2)), s // (2(i + 2)))	57	+ torch.rand(1, 1, s // (2 (i + 2)), s // (2 (i + 2)))
58	for i in range(len(self.anchor_generators))	58	for i in range(len(self.anchor_generators))
59	]	59	]
60	cls_scores, bbox_preds = self.forward(feat)	60	cls_scores, bbox_preds = self.forward(feat)
	@@ -330,14 +330,14 @@ def _demodata_refine_boxes(n_roi, n_img, rng=0):
330	roi_boxes = random_boxes(n_roi, scale=scale, rng=rng)	330	roi_boxes = random_boxes(n_roi, scale=scale, rng=rng)
331	if n_img == 0:	331	if n_img == 0:
332	assert n_roi == 0, 'cannot have any rois if there are no images'	332	assert n_roi == 0, 'cannot have any rois if there are no images'
333	- img_ids = torch.empty((0, ), dtype=torch.long)	333	+ img_ids = torch.empty((0,), dtype=torch.long)
334	roi_boxes = torch.empty((0, 4), dtype=torch.float32)	334	roi_boxes = torch.empty((0, 4), dtype=torch.float32)
335	else:	335	else:
336	- img_ids = rng.randint(0, n_img, (n_roi, ))	336	+ img_ids = rng.randint(0, n_img, (n_roi,))
337	img_ids = torch.from_numpy(img_ids)	337	img_ids = torch.from_numpy(img_ids)
338	rois = torch.cat([img_ids[:, None].float(), roi_boxes], dim=1)	338	rois = torch.cat([img_ids[:, None].float(), roi_boxes], dim=1)
339	# Create other args	339	# Create other args
340	- labels = rng.randint(0, 2, (n_roi, ))	340	+ labels = rng.randint(0, 2, (n_roi,))
341	labels = torch.from_numpy(labels).long()	341	labels = torch.from_numpy(labels).long()
342	bbox_preds = random_boxes(n_roi, scale=scale, rng=rng)	342	bbox_preds = random_boxes(n_roi, scale=scale, rng=rng)
343	# For each image, pretend random positive boxes are gts	343	# For each image, pretend random positive boxes are gts
	@@ -346,7 +346,7 @@ def _demodata_refine_boxes(n_roi, n_img, rng=0):
346	pos_per_img = [sum(lbl_per_img.get(gid, [])) for gid in range(n_img)]	346	pos_per_img = [sum(lbl_per_img.get(gid, [])) for gid in range(n_img)]
347	# randomly generate with numpy then sort with torch	347	# randomly generate with numpy then sort with torch
348	_pos_is_gts = [	348	_pos_is_gts = [
349	- rng.randint(0, 2, (npos, )).astype(np.uint8) for npos in pos_per_img	349	+ rng.randint(0, 2, (npos,)).astype(np.uint8) for npos in pos_per_img
350	]	350	]
351	pos_is_gts = [	351	pos_is_gts = [
352	torch.from_numpy(p).sort(descending=True)[0] for p in _pos_is_gts	352	torch.from_numpy(p).sort(descending=True)[0] for p in _pos_is_gts

MPyTorch/contrib/cv/detection/SOLOv2/tests/test_sampler.py+3-5

	@@ -124,7 +124,6 @@ def _context_for_ohem():
124		124
125		125
126	def test_ohem_sampler():	126	def test_ohem_sampler():
127	-
128	assigner = MaxIoUAssigner(	127	assigner = MaxIoUAssigner(
129	pos_iou_thr=0.5,	128	pos_iou_thr=0.5,
130	neg_iou_thr=0.5,	129	neg_iou_thr=0.5,
	@@ -160,7 +159,7 @@ def test_ohem_sampler():
160	neg_pos_ub=-1,	159	neg_pos_ub=-1,
161	add_gt_as_proposals=True)	160	add_gt_as_proposals=True)
162		161
163	- feats = [torch.rand(1, 256, int(2i), int(2i)) for i in [6, 5, 4, 3, 2]]	162	+ feats = [torch.rand(1, 256, int(2 i), int(2 i)) for i in [6, 5, 4, 3, 2]]
164	sample_result = sampler.sample(	163	sample_result = sampler.sample(
165	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)	164	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)
166		165
	@@ -169,7 +168,6 @@ def test_ohem_sampler():
169		168
170		169
171	def test_ohem_sampler_empty_gt():	170	def test_ohem_sampler_empty_gt():
172	-
173	assigner = MaxIoUAssigner(	171	assigner = MaxIoUAssigner(
174	pos_iou_thr=0.5,	172	pos_iou_thr=0.5,
175	neg_iou_thr=0.5,	173	neg_iou_thr=0.5,
	@@ -200,7 +198,7 @@ def test_ohem_sampler_empty_gt():
200	neg_pos_ub=-1,	198	neg_pos_ub=-1,
201	add_gt_as_proposals=True)	199	add_gt_as_proposals=True)
202		200
203	- feats = [torch.rand(1, 256, int(2i), int(2i)) for i in [6, 5, 4, 3, 2]]	201	+ feats = [torch.rand(1, 256, int(2 i), int(2 i)) for i in [6, 5, 4, 3, 2]]
204		202
205	sample_result = sampler.sample(	203	sample_result = sampler.sample(
206	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)	204	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)
	@@ -240,7 +238,7 @@ def test_ohem_sampler_empty_pred():
240	neg_pos_ub=-1,	238	neg_pos_ub=-1,
241	add_gt_as_proposals=True)	239	add_gt_as_proposals=True)
242		240
243	- feats = [torch.rand(1, 256, int(2i), int(2i)) for i in [6, 5, 4, 3, 2]]	241	+ feats = [torch.rand(1, 256, int(2 i), int(2 i)) for i in [6, 5, 4, 3, 2]]
244		242
245	sample_result = sampler.sample(	243	sample_result = sampler.sample(
246	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)	244	assign_result, bboxes, gt_bboxes, gt_labels, feats=feats)

MPyTorch/contrib/cv/detection/SOLOv2/tools/analyze_logs.py+1-1

	@@ -146,7 +146,7 @@ def add_time_parser(subparsers):
146	'--include-outliers',	146	'--include-outliers',
147	action='store_true',	147	action='store_true',
148	help='include the first value of every epoch when computing '	148	help='include the first value of every epoch when computing '
149	- 'the average time')	149	+ 'the average time')
150		150
151		151
152	def parse_args():	152	def parse_args():

MPyTorch/contrib/cv/detection/SOLOv2/tools/convert_datasets/pascal_voc.py+2-2

	@@ -54,13 +54,13 @@ def parse_xml(args):
54	labels.append(label)	54	labels.append(label)
55	if not bboxes:	55	if not bboxes:
56	bboxes = np.zeros((0, 4))	56	bboxes = np.zeros((0, 4))
57	- labels = np.zeros((0, ))	57	+ labels = np.zeros((0,))
58	else:	58	else:
59	bboxes = np.array(bboxes, ndmin=2) - 1	59	bboxes = np.array(bboxes, ndmin=2) - 1
60	labels = np.array(labels)	60	labels = np.array(labels)
61	if not bboxes_ignore:	61	if not bboxes_ignore:
62	bboxes_ignore = np.zeros((0, 4))	62	bboxes_ignore = np.zeros((0, 4))
63	- labels_ignore = np.zeros((0, ))	63	+ labels_ignore = np.zeros((0,))
64	else:	64	else:
65	bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1	65	bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1
66	labels_ignore = np.array(labels_ignore)	66	labels_ignore = np.array(labels_ignore)

MPyTorch/contrib/cv/detection/SOLOv2/tools/dist_test.sh+1-1

	@@ -14,7 +14,7 @@
14		14
15	#!/usr/bin/env bash	15	#!/usr/bin/env bash
16		16
17	-PYTHON=${PYTHON:-"python"}	17	+PYTHON=${PYTHON:-"python3.7"}
18		18
19	CONFIG=$1	19	CONFIG=$1
20	CHECKPOINT=$2	20	CHECKPOINT=$2

MPyTorch/contrib/cv/detection/SOLOv2/tools/get_flops.py+2-3

	@@ -34,13 +34,12 @@ def parse_args():
34		34
35		35
36	def main():	36	def main():
37	-
38	args = parse_args()	37	args = parse_args()
39		38
40	if len(args.shape) == 1:	39	if len(args.shape) == 1:
41	input_shape = (3, args.shape[0], args.shape[0])	40	input_shape = (3, args.shape[0], args.shape[0])
42	elif len(args.shape) == 2:	41	elif len(args.shape) == 2:
43	- input_shape = (3, ) + tuple(args.shape)	42	+ input_shape = (3,) + tuple(args.shape)
44	else:	43	else:
45	raise ValueError('invalid input shape')	44	raise ValueError('invalid input shape')
46		45
	@@ -54,7 +53,7 @@ def main():
54	else:	53	else:
55	raise NotImplementedError(	54	raise NotImplementedError(
56	'FLOPs counter is currently not currently supported with {}'.	55	'FLOPs counter is currently not currently supported with {}'.
57	- format(model.__class__.__name__))	56	+ format(model.__class__.__name__))
58		57
59	flops, params = get_model_complexity_info(model, input_shape)	58	flops, params = get_model_complexity_info(model, input_shape)
60	split_line = '=' * 30	59	split_line = '=' * 30

MPyTorch/contrib/cv/detection/SOLOv2/tools/robustness_eval.py+1-5

	@@ -20,7 +20,6 @@ import numpy as np
20		20
21		21
22	def print_coco_results(results):	22	def print_coco_results(results):
23	-
24	def _print(result, ap=1, iouThr=None, areaRng='all', maxDets=100):	23	def _print(result, ap=1, iouThr=None, areaRng='all', maxDets=100):
25	iStr = ' {:<18} {} @[ IoU={:<9} \| \	24	iStr = ' {:<18} {} @[ IoU={:<9} \| \
26	area={:>6s} \| maxDets={:>3d} ] = {:0.3f}'	25	area={:>6s} \| maxDets={:>3d} ] = {:0.3f}'
	@@ -31,7 +30,7 @@ def print_coco_results(results):
31	if iouThr is None else '{:0.2f}'.format(iouThr)	30	if iouThr is None else '{:0.2f}'.format(iouThr)
32	print(iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets, result))	31	print(iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets, result))
33		32
34	- stats = np.zeros((12, ))	33	+ stats = np.zeros((12,))
35	stats[0] = _print(results[0], 1)	34	stats[0] = _print(results[0], 1)
36	stats[1] = _print(results[1], 1, iouThr=.5)	35	stats[1] = _print(results[1], 1, iouThr=.5)
37	stats[2] = _print(results[2], 1, iouThr=.75)	36	stats[2] = _print(results[2], 1, iouThr=.75)
	@@ -51,7 +50,6 @@ def get_coco_style_results(filename,
51	metric=None,	50	metric=None,
52	prints='mPC',	51	prints='mPC',
53	aggregate='benchmark'):	52	aggregate='benchmark'):
54	-
55	assert aggregate in ['benchmark', 'all']	53	assert aggregate in ['benchmark', 'all']
56		54
57	if prints == 'all':	55	if prints == 'all':
	@@ -127,7 +125,6 @@ def get_coco_style_results(filename,
127		125
128		126
129	def get_voc_style_results(filename, prints='mPC', aggregate='benchmark'):	127	def get_voc_style_results(filename, prints='mPC', aggregate='benchmark'):
130	-
131	assert aggregate in ['benchmark', 'all']	128	assert aggregate in ['benchmark', 'all']
132		129
133	if prints == 'all':	130	if prints == 'all':
	@@ -201,7 +198,6 @@ def get_results(filename,
201		198
202		199
203	def get_distortions_from_file(filename):	200	def get_distortions_from_file(filename):
204	-
205	eval_output = mmcv.load(filename)	201	eval_output = mmcv.load(filename)
206		202
207	return get_distortions_from_results(eval_output)	203	return get_distortions_from_results(eval_output)

MPyTorch/contrib/cv/detection/SOLOv2/tools/test.py+1-1

	@@ -98,7 +98,7 @@ def collect_results_cpu(result_part, size, tmpdir=None):
98	if tmpdir is None:	98	if tmpdir is None:
99	MAX_LEN = 512	99	MAX_LEN = 512
100	# 32 is whitespace	100	# 32 is whitespace
101	- dir_tensor = torch.full((MAX_LEN, ),	101	+ dir_tensor = torch.full((MAX_LEN,),
102	32,	102	32,
103	dtype=torch.uint8,	103	dtype=torch.uint8,
104	device='cuda')	104	device='cuda')

MPyTorch/contrib/cv/detection/SOLOv2/tools/test_ins.py+125-17

	@@ -15,15 +15,23 @@
15	import argparse	15	import argparse
16	import os	16	import os
17	import os.path as osp	17	import os.path as osp
		18	+import pickle
18	import shutil	19	import shutil
19	import tempfile	20	import tempfile
20		21
21	import mmcv	22	import mmcv
		23	+import apex
		24	+import time
		25	+from apex import amp
22	import torch	26	import torch
		27	+
		28	+if torch.__version__ >= '1.8.1':
		29	+ import torch_npu
23	import torch.nn.functional as F	30	import torch.nn.functional as F
24	import torch.distributed as dist	31	import torch.distributed as dist
25	from mmcv.parallel import MMDataParallel, MMDistributedDataParallel	32	from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
26	from mmcv.runner import init_dist, get_dist_info, load_checkpoint	33	from mmcv.runner import init_dist, get_dist_info, load_checkpoint
		34	+from mmcv.runner import DistSamplerSeedHook, Runner, obj_from_dict
27		35
28	from mmdet.core import coco_eval, results2json, results2json_segm, wrap_fp16_model, tensor2imgs, get_classes	36	from mmdet.core import coco_eval, results2json, results2json_segm, wrap_fp16_model, tensor2imgs, get_classes
29	from mmdet.datasets import build_dataloader, build_dataset	37	from mmdet.datasets import build_dataloader, build_dataset
	@@ -65,7 +73,7 @@ def single_gpu_test(model, data_loader, show=False, verbose=True):
65	seg_result = model(return_loss=False, rescale=not show, **data)	73	seg_result = model(return_loss=False, rescale=not show, **data)
66	result = get_masks(seg_result, num_classes=num_classes)	74	result = get_masks(seg_result, num_classes=num_classes)
67	results.append(result)	75	results.append(result)
68	-	76	+
69	batch_size = data['img'][0].size(0)	77	batch_size = data['img'][0].size(0)
70	for _ in range(batch_size):	78	for _ in range(batch_size):
71	prog_bar.update()	79	prog_bar.update()
	@@ -80,10 +88,13 @@ def multi_gpu_test(model, data_loader, tmpdir=None):
80		88
81	rank, world_size = get_dist_info()	89	rank, world_size = get_dist_info()
82	if rank == 0:	90	if rank == 0:
83	- prog_bar = mmcv.ProgressBar(len(dataset))	91	+ prog_bar = mmcv.ProgressBar(16)
84	for i, data in enumerate(data_loader):	92	for i, data in enumerate(data_loader):
		93	+ if i >= 2:
		94	+ break
85	with torch.no_grad():	95	with torch.no_grad():
86	seg_result = model(return_loss=False, rescale=True, **data)	96	seg_result = model(return_loss=False, rescale=True, **data)
		97	+ torch.npu.synchronize()
87	result = get_masks(seg_result, num_classes=num_classes)	98	result = get_masks(seg_result, num_classes=num_classes)
88	results.append(result)	99	results.append(result)
89		100
	@@ -93,7 +104,7 @@ def multi_gpu_test(model, data_loader, tmpdir=None):
93	prog_bar.update()	104	prog_bar.update()
94		105
95	# collect results from all ranks	106	# collect results from all ranks
96	- results = collect_results(results, len(dataset), tmpdir)	107	+ results = collect_results(results, 16, tmpdir)
97		108
98	return results	109	return results
99		110
	@@ -104,14 +115,14 @@ def collect_results(result_part, size, tmpdir=None):
104	if tmpdir is None:	115	if tmpdir is None:
105	MAX_LEN = 512	116	MAX_LEN = 512
106	# 32 is whitespace	117	# 32 is whitespace
107	- dir_tensor = torch.full((MAX_LEN, ),	118	+ dir_tensor = torch.full((MAX_LEN,),
108	32,	119	32,
109	dtype=torch.uint8,	120	dtype=torch.uint8,
110	- device='cuda')	121	+ device='npu')
111	if rank == 0:	122	if rank == 0:
112	tmpdir = tempfile.mkdtemp()	123	tmpdir = tempfile.mkdtemp()
113	tmpdir = torch.tensor(	124	tmpdir = torch.tensor(
114	- bytearray(tmpdir.encode()), dtype=torch.uint8, device='cuda')	125	+ bytearray(tmpdir.encode()), dtype=torch.uint8, device='npu')
115	dir_tensor[:len(tmpdir)] = tmpdir	126	dir_tensor[:len(tmpdir)] = tmpdir
116	dist.broadcast(dir_tensor, 0)	127	dist.broadcast(dir_tensor, 0)
117	tmpdir = dir_tensor.cpu().numpy().tobytes().decode().rstrip()	128	tmpdir = dir_tensor.cpu().numpy().tobytes().decode().rstrip()
	@@ -179,19 +190,104 @@ def parse_args():
179	return args	190	return args
180		191
181		192
		193	+def build_optimizer(model, optimizer_cfg):
		194	+ """Build optimizer from configs.
		195	+
		196	+ Args:
		197	+ model (:obj:`nn.Module`): The model with parameters to be optimized.
		198	+ optimizer_cfg (dict): The config dict of the optimizer.
		199	+ Positional fields are:
		200	+ - type: class name of the optimizer.
		201	+ - lr: base learning rate.
		202	+ Optional fields are:
		203	+ - any arguments of the corresponding optimizer type, e.g.,
		204	+ weight_decay, momentum, etc.
		205	+ - paramwise_options: a dict with 3 accepted fileds
		206	+ (bias_lr_mult, bias_decay_mult, norm_decay_mult).
		207	+ `bias_lr_mult` and `bias_decay_mult` will be multiplied to
		208	+ the lr and weight decay respectively for all bias parameters
		209	+ (except for the normalization layers), and
		210	+ `norm_decay_mult` will be multiplied to the weight decay
		211	+ for all weight and bias parameters of normalization layers.
		212	+
		213	+ Returns:
		214	+ torch.optim.Optimizer: The initialized optimizer.
		215	+
		216	+ Example:
		217	+ >>> model = torch.nn.modules.Conv1d(1, 1, 1)
		218	+ >>> optimizer_cfg = dict(type='SGD', lr=0.01, momentum=0.9,
		219	+ >>> weight_decay=0.0001)
		220	+ >>> optimizer = build_optimizer(model, optimizer_cfg)
		221	+ """
		222	+ if hasattr(model, 'module'):
		223	+ model = model.module
		224	+
		225	+ optimizer_cfg = optimizer_cfg.copy()
		226	+ paramwise_options = optimizer_cfg.pop('paramwise_options', None)
		227	+ # if no paramwise option is specified, just use the global setting
		228	+ if paramwise_options is None:
		229	+ return obj_from_dict(optimizer_cfg, torch.optim,
		230	+ dict(params=model.parameters()))
		231	+ else:
		232	+ assert isinstance(paramwise_options, dict)
		233	+ # get base lr and weight decay
		234	+ base_lr = optimizer_cfg['lr']
		235	+ base_wd = optimizer_cfg.get('weight_decay', None)
		236	+ # weight_decay must be explicitly specified if mult is specified
		237	+ if ('bias_decay_mult' in paramwise_options
		238	+ or 'norm_decay_mult' in paramwise_options):
		239	+ assert base_wd is not None
		240	+ # get param-wise options
		241	+ bias_lr_mult = paramwise_options.get('bias_lr_mult', 1.)
		242	+ bias_decay_mult = paramwise_options.get('bias_decay_mult', 1.)
		243	+ norm_decay_mult = paramwise_options.get('norm_decay_mult', 1.)
		244	+ # set param-wise lr and weight decay
		245	+ params = []
		246	+ for name, param in model.named_parameters():
		247	+ param_group = {'params': [param]}
		248	+ if not param.requires_grad:
		249	+ # FP16 training needs to copy gradient/weight between master
		250	+ # weight copy and model weight, it is convenient to keep all
		251	+ # parameters here to align with model.parameters()
		252	+ params.append(param_group)
		253	+ continue
		254	+
		255	+ # for norm layers, overwrite the weight decay of weight and bias
		256	+ # TODO: obtain the norm layer prefixes dynamically
		257	+ if re.search(r'(bn\|gn)(\d+)?.(weight\|bias)', name):
		258	+ if base_wd is not None:
		259	+ param_group['weight_decay'] = base_wd * norm_decay_mult
		260	+ # for other layers, overwrite both lr and weight decay of bias
		261	+ elif name.endswith('.bias'):
		262	+ param_group['lr'] = base_lr * bias_lr_mult
		263	+ if base_wd is not None:
		264	+ param_group['weight_decay'] = base_wd * bias_decay_mult
		265	+ # otherwise use the global settings
		266	+
		267	+ params.append(param_group)
		268	+
		269	+ optimizer_cls = getattr(torch.optim, optimizer_cfg.pop('type'))
		270	+ return optimizer_cls(params, **optimizer_cfg)
		271	+
		272	+
182	def main():	273	def main():
183	args = parse_args()	274	args = parse_args()
		275	+ option = {}
		276	+ option["ACL_OP_COMPILER_CACHE_MODE"] = 'enable'
		277	+ option["ACL_OP_COMPILER_CACHE_DIR"] = './cache'
184		278
		279	+ option["ACL_OP_SELECT_IMPL_MODE"] = 'high_precision'
		280	+ option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'Sqrt'
		281	+ print('option', option)
		282	+ torch.npu.set_option(option)
185	assert args.out or args.show or args.json_out, \	283	assert args.out or args.show or args.json_out, \
186	('Please specify at least one operation (save or show the results) '	284	('Please specify at least one operation (save or show the results) '
187	'with the argument "--out" or "--show" or "--json_out"')	285	'with the argument "--out" or "--show" or "--json_out"')
188	-
189	if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):	286	if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
190	raise ValueError('The output file must be a pkl file.')	287	raise ValueError('The output file must be a pkl file.')
191		288
192	if args.json_out is not None and args.json_out.endswith('.json'):	289	if args.json_out is not None and args.json_out.endswith('.json'):
193	args.json_out = args.json_out[:-5]	290	args.json_out = args.json_out[:-5]
194	-
195	cfg = mmcv.Config.fromfile(args.config)	291	cfg = mmcv.Config.fromfile(args.config)
196	if args.data_root:	292	if args.data_root:
197	cfg.data_root = args.data_root	293	cfg.data_root = args.data_root
	@@ -205,12 +301,15 @@ def main():
205	if args.gpu_ids is not None:	301	if args.gpu_ids is not None:
206	torch.npu.set_device(args.gpu_ids[0])	302	torch.npu.set_device(args.gpu_ids[0])
207	# init distributed env first, since logger depends on the dist info.	303	# init distributed env first, since logger depends on the dist info.
208	- if args.launcher == 'none':	304	+ if "WORLD_SIZE" in os.environ and int(os.environ["WORLD_SIZE"]) > 1:
209	- distributed = False
210	- else:
211	distributed = True	305	distributed = True
212	- init_dist(args.launcher, **cfg.dist_params)	306	+ rank = int(os.environ['RANK'])
213	-	307	+ num_gpus = torch.npu.device_count()
		308	+ torch.npu.set_device(rank % num_gpus)
		309	+ torch.distributed.init_process_group(backend='hccl', init_method='env://',
		310	+ world_size=int(os.environ['WORLD_SIZE']), rank=args.local_rank)
		311	+ else:
		312	+ distributed = False
214	# build the dataloader	313	# build the dataloader
215	# TODO: support multiple images per gpu (only minor changes are needed)	314	# TODO: support multiple images per gpu (only minor changes are needed)
216	print('arg data root', cfg.data_root)	315	print('arg data root', cfg.data_root)
	@@ -225,10 +324,11 @@ def main():
225	# build the model and load checkpoint	324	# build the model and load checkpoint
226	model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)	325	model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
227	fp16_cfg = cfg.get('fp16', None)	326	fp16_cfg = cfg.get('fp16', None)
228	- if fp16_cfg is not None:	327	+ # if fp16_cfg is not None:
229	- wrap_fp16_model(model)	328	+ # wrap_fp16_model(model)
230		329
231	while not osp.isfile(args.checkpoint):	330	while not osp.isfile(args.checkpoint):
		331	+ print(args.checkpoint)
232	print('Waiting for {} to exist...'.format(args.checkpoint))	332	print('Waiting for {} to exist...'.format(args.checkpoint))
233	time.sleep(60)	333	time.sleep(60)
234		334
	@@ -241,10 +341,18 @@ def main():
241	model.CLASSES = dataset.CLASSES	341	model.CLASSES = dataset.CLASSES
242		342
243	if not distributed:	343	if not distributed:
244	- model = MMDataParallel(model.npu(), device_ids=[0])	344	+ amp.register_float_function(torch, 'sigmoid')
		345	+ optimizer = build_optimizer(model, cfg.optimizer)
		346	+ model, optimizer = amp.initialize(model.npu(), optimizer,
		347	+ opt_level='O1', loss_scale=128.0, combine_grad=True)
		348	+ model = MMDataParallel(model, device_ids=[0])
245	outputs = single_gpu_test(model, data_loader)	349	outputs = single_gpu_test(model, data_loader)
246	else:	350	else:
247	- model = MMDistributedDataParallel(model.npu())	351	+ amp.register_float_function(torch, 'sigmoid')
		352	+ optimizer = build_optimizer(model, cfg.optimizer)
		353	+ model, optimizer = amp.initialize(model.npu(), optimizer,
		354	+ opt_level='O1', loss_scale=128.0, combine_grad=True)
		355	+ model = MMDistributedDataParallel(model)
248	outputs = multi_gpu_test(model, data_loader, args.tmpdir)	356	outputs = multi_gpu_test(model, data_loader, args.tmpdir)
249		357
250	rank, _ = get_dist_info()	358	rank, _ = get_dist_info()

MPyTorch/contrib/cv/detection/SOLOv2/tools/test_ins_vis.py+11-9

	@@ -32,6 +32,7 @@ import cv2
32	import numpy as np	32	import numpy as np
33	import matplotlib.cm as cm	33	import matplotlib.cm as cm
34		34
		35	+
35	def vis_seg(data, result, img_norm_cfg, data_id, colors, score_thr, save_dir):	36	def vis_seg(data, result, img_norm_cfg, data_id, colors, score_thr, save_dir):
36	img_tensor = data['img'][0]	37	img_tensor = data['img'][0]
37	img_metas = data['img_meta'][0].data[0]	38	img_metas = data['img_meta'][0].data[0]
	@@ -44,7 +45,7 @@ def vis_seg(data, result, img_norm_cfg, data_id, colors, score_thr, save_dir):
44	continue	45	continue
45	h, w, _ = img_meta['img_shape']	46	h, w, _ = img_meta['img_shape']
46	img_show = img[:h, :w, :]	47	img_show = img[:h, :w, :]
47	-	48	+
48	seg_label = cur_result[0]	49	seg_label = cur_result[0]
49	seg_label = seg_label.cpu().numpy().astype(np.uint8)	50	seg_label = seg_label.cpu().numpy().astype(np.uint8)
50	cate_label = cur_result[1]	51	cate_label = cur_result[1]
	@@ -70,12 +71,12 @@ def vis_seg(data, result, img_norm_cfg, data_id, colors, score_thr, save_dir):
70		71
71	seg_show = img_show.copy()	72	seg_show = img_show.copy()
72	for idx in range(num_mask):	73	for idx in range(num_mask):
73	- idx = -(idx+1)	74	+ idx = -(idx + 1)
74	- cur_mask = seg_label[idx, :,:]	75	+ cur_mask = seg_label[idx, :, :]
75	cur_mask = mmcv.imresize(cur_mask, (w, h))	76	cur_mask = mmcv.imresize(cur_mask, (w, h))
76	cur_mask = (cur_mask > 0.5).astype(np.uint8)	77	cur_mask = (cur_mask > 0.5).astype(np.uint8)
77	if cur_mask.sum() == 0:	78	if cur_mask.sum() == 0:
78	- continue	79	+ continue
79	color_mask = np.random.randint(	80	color_mask = np.random.randint(
80	0, 256, (1, 3), dtype=np.uint8)	81	0, 256, (1, 3), dtype=np.uint8)
81	cur_mask_bool = cur_mask.astype(np.bool)	82	cur_mask_bool = cur_mask.astype(np.bool)
	@@ -85,7 +86,7 @@ def vis_seg(data, result, img_norm_cfg, data_id, colors, score_thr, save_dir):
85	cur_score = cate_score[idx]	86	cur_score = cate_score[idx]
86		87
87	label_text = class_names[cur_cate]	88	label_text = class_names[cur_cate]
88	- #label_text += '\|{:.02f}'.format(cur_score)	89	+ # label_text += '\|{:.02f}'.format(cur_score)
89	# center	90	# center
90	center_y, center_x = ndimage.measurements.center_of_mass(cur_mask)	91	center_y, center_x = ndimage.measurements.center_of_mass(cur_mask)
91	vis_pos = (max(int(center_x) - 10, 0), int(center_y))	92	vis_pos = (max(int(center_x) - 10, 0), int(center_y))
	@@ -99,8 +100,8 @@ def single_gpu_test(model, data_loader, args, cfg=None, verbose=True):
99	results = []	100	results = []
100	dataset = data_loader.dataset	101	dataset = data_loader.dataset
101		102
102	- class_num = 1000 # ins	103	+ class_num = 1000 # ins
103	- colors = [(np.random.random((1, 3)) * 255).tolist()[0] for i in range(class_num)]	104	+ colors = [(np.random.random((1, 3)) * 255).tolist()[0] for i in range(class_num)]
104		105
105	prog_bar = mmcv.ProgressBar(len(dataset))	106	prog_bar = mmcv.ProgressBar(len(dataset))
106	for i, data in enumerate(data_loader):	107	for i, data in enumerate(data_loader):
	@@ -110,7 +111,8 @@ def single_gpu_test(model, data_loader, args, cfg=None, verbose=True):
110	results.append(result)	111	results.append(result)
111		112
112	if verbose:	113	if verbose:
113	- vis_seg(data, seg_result, cfg.img_norm_cfg, data_id=i, colors=colors, score_thr=args.score_thr, save_dir=args.save_dir)	114	+ vis_seg(data, seg_result, cfg.img_norm_cfg, data_id=i, colors=colors, score_thr=args.score_thr,
		115	+ save_dir=args.save_dir)
114		116
115	batch_size = data['img'][0].size(0)	117	batch_size = data['img'][0].size(0)
116	for _ in range(batch_size):	118	for _ in range(batch_size):
	@@ -147,7 +149,7 @@ def collect_results(result_part, size, tmpdir=None):
147	if tmpdir is None:	149	if tmpdir is None:
148	MAX_LEN = 512	150	MAX_LEN = 512
149	# 32 is whitespace	151	# 32 is whitespace
150	- dir_tensor = torch.full((MAX_LEN, ),	152	+ dir_tensor = torch.full((MAX_LEN,),
151	32,	153	32,
152	dtype=torch.uint8,	154	dtype=torch.uint8,
153	device='cuda')	155	device='cuda')

MPyTorch/contrib/cv/detection/SOLOv2/tools/test_robustness.py+3-3

	@@ -162,7 +162,7 @@ def collect_results(result_part, size, tmpdir=None):
162	if tmpdir is None:	162	if tmpdir is None:
163	MAX_LEN = 512	163	MAX_LEN = 512
164	# 32 is whitespace	164	# 32 is whitespace
165	- dir_tensor = torch.full((MAX_LEN, ),	165	+ dir_tensor = torch.full((MAX_LEN,),
166	32,	166	32,
167	dtype=torch.uint8,	167	dtype=torch.uint8,
168	device='cuda')	168	device='cuda')
	@@ -395,8 +395,8 @@ def main():
395	rank, _ = get_dist_info()	395	rank, _ = get_dist_info()
396	if args.out and rank == 0:	396	if args.out and rank == 0:
397	eval_results_filename = (	397	eval_results_filename = (
398	- osp.splitext(args.out)[0] + '_results' +	398	+ osp.splitext(args.out)[0] + '_results' +
399	- osp.splitext(args.out)[1])	399	+ osp.splitext(args.out)[1])
400	mmcv.dump(outputs, args.out)	400	mmcv.dump(outputs, args.out)
401	eval_types = args.eval	401	eval_types = args.eval
402	if cfg.dataset_type == 'VOCDataset':	402	if cfg.dataset_type == 'VOCDataset':

MPyTorch/contrib/cv/detection/SOLOv2/tools/train.py+17-3

	@@ -20,6 +20,9 @@ import time
20		20
21	import mmcv	21	import mmcv
22	import torch	22	import torch
		23	+
		24	+if torch.__version__ >= '1.8.1':
		25	+ import torch_npu
23	from mmcv import Config	26	from mmcv import Config
24	from mmcv.runner import init_dist, load_state_dict	27	from mmcv.runner import init_dist, load_state_dict
25		28
	@@ -29,6 +32,7 @@ from mmdet.datasets import build_dataset
29	from mmdet.models import build_detector	32	from mmdet.models import build_detector
30	from mmdet.utils import get_root_logger	33	from mmdet.utils import get_root_logger
31		34
		35	+
32	def parse_args():	36	def parse_args():
33	parser = argparse.ArgumentParser(description='Train a detector')	37	parser = argparse.ArgumentParser(description='Train a detector')
34	parser.add_argument('config', help='train config file path')	38	parser.add_argument('config', help='train config file path')
	@@ -44,7 +48,7 @@ def parse_args():
44	type=int,	48	type=int,
45	default=1,	49	default=1,
46	help='number of gpus to use '	50	help='number of gpus to use '
47	- '(only applicable to non-distributed training)')	51	+ '(only applicable to non-distributed training)')
48	parser.add_argument(	52	parser.add_argument(
49	'--data_root',	53	'--data_root',
50	help='the path of dataset',	54	help='the path of dataset',
	@@ -61,6 +65,7 @@ def parse_args():
61	action='store_true',	65	action='store_true',
62	help='whether fine-tune model, change class num + 1')	66	help='whether fine-tune model, change class num + 1')
63	parser.add_argument('--total_epochs', type=int, default=12, help='random seed')	67	parser.add_argument('--total_epochs', type=int, default=12, help='random seed')
		68	+ parser.add_argument('--train_performance', type=bool, default=False, help='train performace')
64	parser.add_argument(	69	parser.add_argument(
65	'--deterministic',	70	'--deterministic',
66	action='store_true',	71	action='store_true',
	@@ -88,6 +93,14 @@ def parse_args():
88		93
89		94
90	def main():	95	def main():
		96	+ option = {}
		97	+ option["ACL_OP_COMPILER_CACHE_MODE"] = 'enable'
		98	+ option["ACL_OP_COMPILER_CACHE_DIR"] = './cache'
		99	+
		100	+ option["ACL_OP_SELECT_IMPL_MODE"] = 'high_precision'
		101	+ option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'Sqrt'
		102	+ print('option', option)
		103	+ torch.npu.set_option(option)
91	os.environ['MASTER_ADDR'] = '127.0.0.1'	104	os.environ['MASTER_ADDR'] = '127.0.0.1'
92	os.environ['MASTER_PORT'] = '29688'	105	os.environ['MASTER_PORT'] = '29688'
93	args = parse_args()	106	args = parse_args()
	@@ -177,8 +190,9 @@ def main():
177	cfg,	190	cfg,
178	distributed=distributed,	191	distributed=distributed,
179	validate=args.validate,	192	validate=args.validate,
180	- timestamp=timestamp)	193	+ timestamp=timestamp,
		194	+ train_performance=args.train_performance)
181		195
182		196
183	if __name__ == '__main__':	197	if __name__ == '__main__':
184	- main()	198	+ main()

	@@ -74,8 +74,8 @@ class TestImage(object):
74	in_img = np.random.rand(10, 10, 3).astype(np.float32)	74	in_img = np.random.rand(10, 10, 3).astype(np.float32)
75	out_img = mmcv.bgr2gray(in_img)	75	out_img = mmcv.bgr2gray(in_img)
76	computed_gray = (	76	computed_gray = (
77	- in_img[:, :, 0] * 0.114 + in_img[:, :, 1] * 0.587 +	77	+ in_img[:, :, 0] * 0.114 + in_img[:, :, 1] * 0.587 +
78	- in_img[:, :, 2] * 0.299)	78	+ in_img[:, :, 2] * 0.299)
79	assert_array_almost_equal(out_img, computed_gray, decimal=4)	79	assert_array_almost_equal(out_img, computed_gray, decimal=4)
80	out_img_3d = mmcv.bgr2gray(in_img, True)	80	out_img_3d = mmcv.bgr2gray(in_img, True)
81	assert out_img_3d.shape == (10, 10, 1)	81	assert out_img_3d.shape == (10, 10, 1)
	@@ -85,8 +85,8 @@ class TestImage(object):
85	in_img = np.random.rand(10, 10, 3).astype(np.float32)	85	in_img = np.random.rand(10, 10, 3).astype(np.float32)
86	out_img = mmcv.rgb2gray(in_img)	86	out_img = mmcv.rgb2gray(in_img)
87	computed_gray = (	87	computed_gray = (
88	- in_img[:, :, 0] * 0.299 + in_img[:, :, 1] * 0.587 +	88	+ in_img[:, :, 0] * 0.299 + in_img[:, :, 1] * 0.587 +
89	- in_img[:, :, 2] * 0.114)	89	+ in_img[:, :, 2] * 0.114)
90	assert_array_almost_equal(out_img, computed_gray, decimal=4)	90	assert_array_almost_equal(out_img, computed_gray, decimal=4)
91	out_img_3d = mmcv.rgb2gray(in_img, True)	91	out_img_3d = mmcv.rgb2gray(in_img, True)
92	assert out_img_3d.shape == (10, 10, 1)	92	assert out_img_3d.shape == (10, 10, 1)
	@@ -324,7 +324,7 @@ class TestImage(object):
324	(15, 2, 3), dtype='uint8'), padded_img[:, 10:, :])	324	(15, 2, 3), dtype='uint8'), padded_img[:, 10:, :])
325		325
326	with pytest.raises(AssertionError):	326	with pytest.raises(AssertionError):
327	- mmcv.impad(img, (15, ), 0)	327	+ mmcv.impad(img, (15,), 0)
328	with pytest.raises(AssertionError):	328	with pytest.raises(AssertionError):
329	mmcv.impad(img, (5, 5), 0)	329	mmcv.impad(img, (5, 5), 0)
330	with pytest.raises(AssertionError):	330	with pytest.raises(AssertionError):

	@@ -374,7 +374,7 @@ class FCOSHead(nn.Module):
374	gt_bboxes.new_zeros((num_points, 4))	374	gt_bboxes.new_zeros((num_points, 4))
375		375
376	areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) * (	376	areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) * (
377	- gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1)	377	+ gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1)
378	# TODO: figure out why these two are different	378	# TODO: figure out why these two are different
379	# areas = areas[None].expand(num_points, num_gts)	379	# areas = areas[None].expand(num_points, num_gts)
380	areas = areas[None].repeat(num_points, 1)	380	areas = areas[None].repeat(num_points, 1)
	@@ -397,8 +397,8 @@ class FCOSHead(nn.Module):
397	# condition2: limit the regression range for each location	397	# condition2: limit the regression range for each location
398	max_regress_distance = bbox_targets.max(-1)[0]	398	max_regress_distance = bbox_targets.max(-1)[0]
399	inside_regress_range = (	399	inside_regress_range = (
400	- max_regress_distance >= regress_ranges[..., 0]) & (	400	+ max_regress_distance >= regress_ranges[..., 0]) & (
401	- max_regress_distance <= regress_ranges[..., 1])	401	+ max_regress_distance <= regress_ranges[..., 1])
402		402
403	# if there are still more than one objects for a location,	403	# if there are still more than one objects for a location,
404	# we choose the one with minimal area	404	# we choose the one with minimal area
	@@ -417,6 +417,6 @@ class FCOSHead(nn.Module):
417	left_right = pos_bbox_targets[:, [0, 2]]	417	left_right = pos_bbox_targets[:, [0, 2]]
418	top_bottom = pos_bbox_targets[:, [1, 3]]	418	top_bottom = pos_bbox_targets[:, [1, 3]]
419	centerness_targets = (	419	centerness_targets = (
420	- left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0]) * (	420	+ left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0]) * (
421	- top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0])	421	+ top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0])
422	return torch.sqrt(centerness_targets)	422	return torch.sqrt(centerness_targets)

	@@ -187,9 +187,9 @@ class RepPointsHead(nn.Module):
187	"""	187	"""
188	pts_reshape = pts.view(pts.shape[0], -1, 2, *pts.shape[2:])	188	pts_reshape = pts.view(pts.shape[0], -1, 2, *pts.shape[2:])
189	pts_y = pts_reshape[:, :, 0, ...] if y_first else pts_reshape[:, :, 1,	189	pts_y = pts_reshape[:, :, 0, ...] if y_first else pts_reshape[:, :, 1,
190	- ...]	190	+ ...]
191	pts_x = pts_reshape[:, :, 1, ...] if y_first else pts_reshape[:, :, 0,	191	pts_x = pts_reshape[:, :, 1, ...] if y_first else pts_reshape[:, :, 0,
192	- ...]	192	+ ...]
193	if self.transform_method == 'minmax':	193	if self.transform_method == 'minmax':
194	bbox_left = pts_x.min(dim=1, keepdim=True)[0]	194	bbox_left = pts_x.min(dim=1, keepdim=True)[0]
195	bbox_right = pts_x.max(dim=1, keepdim=True)[0]	195	bbox_right = pts_x.max(dim=1, keepdim=True)[0]
	@@ -212,7 +212,7 @@ class RepPointsHead(nn.Module):
212	pts_y_std = torch.std(pts_y - pts_y_mean, dim=1, keepdim=True)	212	pts_y_std = torch.std(pts_y - pts_y_mean, dim=1, keepdim=True)
213	pts_x_std = torch.std(pts_x - pts_x_mean, dim=1, keepdim=True)	213	pts_x_std = torch.std(pts_x - pts_x_mean, dim=1, keepdim=True)
214	moment_transfer = (self.moment_transfer * self.moment_mul) + (	214	moment_transfer = (self.moment_transfer * self.moment_mul) + (
215	- self.moment_transfer.detach() * (1 - self.moment_mul))	215	+ self.moment_transfer.detach() * (1 - self.moment_mul))
216	moment_width_transfer = moment_transfer[0]	216	moment_width_transfer = moment_transfer[0]
217	moment_height_transfer = moment_transfer[1]	217	moment_height_transfer = moment_transfer[1]
218	half_width = pts_x_std * torch.exp(moment_width_transfer)	218	half_width = pts_x_std * torch.exp(moment_width_transfer)
	@@ -221,7 +221,7 @@ class RepPointsHead(nn.Module):
221	pts_x_mean - half_width, pts_y_mean - half_height,	221	pts_x_mean - half_width, pts_y_mean - half_height,
222	pts_x_mean + half_width, pts_y_mean + half_height	222	pts_x_mean + half_width, pts_y_mean + half_height
223	],	223	],
224	- dim=1)	224	+ dim=1)
225	else:	225	else:
226	raise NotImplementedError	226	raise NotImplementedError
227	return bbox	227	return bbox