[Feature] Add ViPNAS mbv3

configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_coco.md

@@ -36,4 +36,5 @@ Results on COCO val2017 with detector having human AP of 56.4 on COCO val2017 da
 
 | Arch  | Input Size | AP | AP<sup>50</sup> | AP<sup>75</sup> | AR | AR<sup>50</sup> | ckpt | log |
 | :-------------- | :-----------: | :------: | :------: | :------: | :------: | :------: |:------: |:------: |
-| [S-ViPNAS-Res50](/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_res50_coco_256x192.py)  | 256x192 | 0.711 | 0.893 | 0.789 | 0.769 | 0.769 | [ckpt](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_res50_coco_256x192-cc43b466_20210624.pth) | [log](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_res50_coco_256x192_20210624.log.json) |
+| [S-ViPNAS-MobileNetV3](/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_mbv3_coco_256x192.py)  | 256x192 | 0.700 | 0.887 | 0.778 | 0.757 | 0.929 | [ckpt](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_mbv3_coco_256x192-7018731a_20211122.pth) | [log](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_mbv3_coco_256x192_20211122.log.json) |
+| [S-ViPNAS-Res50](/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_res50_coco_256x192.py)  | 256x192 | 0.711 | 0.893 | 0.789 | 0.769 | 0.934 | [ckpt](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_res50_coco_256x192-cc43b466_20210624.pth) | [log](https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_res50_coco_256x192_20210624.log.json) |

configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_coco.yml

@@ -4,12 +4,29 @@ Collections:
     Title: 'ViPNAS: Efficient Video Pose Estimation via Neural Architecture Search'
     URL: https://arxiv.org/abs/2105.10154
 Models:
-- Config: configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_res50_coco_256x192.py
+- Config: configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_mbv3_coco_256x192.py
   In Collection: ViPNAS
   Metadata:
-    Architecture:
+    Architecture: &id001
     - ViPNAS
     Training Data: COCO
+  Name: topdown_heatmap_vipnas_mbv3_coco_256x192
+  README: configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_coco.md
+  Results:
+  - Dataset: COCO
+    Metrics:
+      AP: 0.7
+      [email protected]: 0.887
+      [email protected]: 0.778
+      AR: 0.757
+      [email protected]: 0.929
+    Task: Body 2D Keypoint
+  Weights: https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_mbv3_coco_256x192-7018731a_20211122.pth
+- Config: configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_res50_coco_256x192.py
+  In Collection: ViPNAS
+  Metadata:
+    Architecture: *id001
+    Training Data: COCO
   Name: topdown_heatmap_vipnas_res50_coco_256x192
   README: configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_coco.md
   Results:
@@ -19,6 +36,6 @@ Models:
       [email protected]: 0.893
       [email protected]: 0.789
       AR: 0.769
-      [email protected]: 0.769
+      [email protected]: 0.934
     Task: Body 2D Keypoint
   Weights: https://download.openmmlab.com/mmpose/top_down/vipnas/vipnas_res50_coco_256x192-cc43b466_20210624.pth

configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/vipnas_mbv3_coco_256x192.py

@@ -0,0 +1,144 @@
+log_level = 'INFO'
+load_from = None
+resume_from = None
+dist_params = dict(backend='nccl')
+workflow = [('train', 1)]
+checkpoint_config = dict(interval=10)
+evaluation = dict(interval=10, metric='mAP', key_indicator='AP')
+
+optimizer = dict(
+    type='Adam',
+    lr=5e-4,
+)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[170, 200])
+total_epochs = 210
+log_config = dict(
+    interval=50,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+
+channel_cfg = dict(
+    num_output_channels=17,
+    dataset_joints=17,
+    dataset_channel=[
+        [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16],
+    ],
+    inference_channel=[
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
+    ])
+
+# model settings
+model = dict(
+    type='TopDown',
+    pretrained=None,
+    backbone=dict(type='ViPNAS_MobileNetV3'),
+    keypoint_head=dict(
+        type='ViPNASHeatmapSimpleHead',
+        in_channels=160,
+        out_channels=channel_cfg['num_output_channels'],
+        num_deconv_filters=(160, 160, 160),
+        num_deconv_groups=(160, 160, 160),
+        loss_keypoint=dict(type='JointsMSELoss', use_target_weight=True)),
+    train_cfg=dict(),
+    test_cfg=dict(
+        flip_test=True,
+        post_process='default',
+        shift_heatmap=True,
+        modulate_kernel=11))
+
+data_cfg = dict(
+    image_size=[192, 256],
+    heatmap_size=[48, 64],
+    num_output_channels=channel_cfg['num_output_channels'],
+    num_joints=channel_cfg['dataset_joints'],
+    dataset_channel=channel_cfg['dataset_channel'],
+    inference_channel=channel_cfg['inference_channel'],
+    soft_nms=False,
+    nms_thr=1.0,
+    oks_thr=0.9,
+    vis_thr=0.2,
+    use_gt_bbox=False,
+    det_bbox_thr=0.0,
+    bbox_file='data/coco/person_detection_results/'
+    'COCO_val2017_detections_AP_H_56_person.json',
+)
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='TopDownRandomFlip', flip_prob=0.5),
+    dict(
+        type='TopDownHalfBodyTransform',
+        num_joints_half_body=8,
+        prob_half_body=0.3),
+    dict(
+        type='TopDownGetRandomScaleRotation', rot_factor=30,
+        scale_factor=0.25),
+    dict(type='TopDownAffine'),
+    dict(type='ToTensor'),
+    dict(
+        type='NormalizeTensor',
+        mean=[0.485, 0.456, 0.406],
+        std=[0.229, 0.224, 0.225]),
+    dict(type='TopDownGenerateTarget', sigma=2),
+    dict(
+        type='Collect',
+        keys=['img', 'target', 'target_weight'],
+        meta_keys=[
+            'image_file', 'joints_3d', 'joints_3d_visible', 'center', 'scale',
+            'rotation', 'bbox_score', 'flip_pairs'
+        ]),
+]
+
+val_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='TopDownAffine'),
+    dict(type='ToTensor'),
+    dict(
+        type='NormalizeTensor',
+        mean=[0.485, 0.456, 0.406],
+        std=[0.229, 0.224, 0.225]),
+    dict(
+        type='Collect',
+        keys=['img'],
+        meta_keys=[
+            'image_file', 'center', 'scale', 'rotation', 'bbox_score',
+            'flip_pairs'
+        ]),
+]
+
+test_pipeline = val_pipeline
+
+data_root = 'data/coco'
+data = dict(
+    samples_per_gpu=64,
+    workers_per_gpu=2,
+    val_dataloader=dict(samples_per_gpu=32),
+    test_dataloader=dict(samples_per_gpu=32),
+    train=dict(
+        type='TopDownCocoDataset',
+        ann_file=f'{data_root}/annotations/person_keypoints_train2017.json',
+        img_prefix=f'{data_root}/train2017/',
+        data_cfg=data_cfg,
+        pipeline=train_pipeline),
+    val=dict(
+        type='TopDownCocoDataset',
+        ann_file=f'{data_root}/annotations/person_keypoints_val2017.json',
+        img_prefix=f'{data_root}/val2017/',
+        data_cfg=data_cfg,
+        pipeline=val_pipeline),
+    test=dict(
+        type='TopDownCocoDataset',
+        ann_file=f'{data_root}/annotations/person_keypoints_val2017.json',
+        img_prefix=f'{data_root}/val2017/',
+        data_cfg=data_cfg,
+        pipeline=val_pipeline),
+)

configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_coco-wholebody.yml

@@ -1,20 +1,17 @@
 Collections:
-- Metadata:
-    Architecture:
-    - ViPNAS
-  Name: wholebody--2d_kpt_sview_rgb_img--topdown_heatmap--coco-wholebody--vipnas_coco-wholebody
-    [ViPNAS@COCO-WholeBody]
+- Name: ViPNAS
   Paper:
     Title: 'ViPNAS: Efficient Video Pose Estimation via Neural Architecture Search'
     URL: https://arxiv.org/abs/2105.10154
-  README: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_coco-wholebody.md
 Models:
 - Config: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_res50_coco_wholebody_256x192.py
-  In Collection: wholebody--2d_kpt_sview_rgb_img--topdown_heatmap--coco-wholebody--vipnas_coco-wholebody
-    [ViPNAS@COCO-WholeBody]
+  In Collection: ViPNAS
   Metadata:
+    Architecture:
+    - ViPNAS
     Training Data: COCO-WholeBody
   Name: topdown_heatmap_vipnas_res50_coco_wholebody_256x192
+  README: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_coco-wholebody.md
   Results:
   - Dataset: COCO-WholeBody
     Metrics:

configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_dark_coco-wholebody.yml

@@ -1,21 +1,18 @@
 Collections:
-- Metadata:
-    Architecture:
-    - ViPNAS
-    - DarkPose
-  Name: wholebody--2d_kpt_sview_rgb_img--topdown_heatmap--coco-wholebody--vipnas_dark_coco-wholebody
-    [ViPNAS+DarkPose@COCO-WholeBody]
+- Name: ViPNAS
   Paper:
     Title: 'ViPNAS: Efficient Video Pose Estimation via Neural Architecture Search'
     URL: https://arxiv.org/abs/2105.10154
-  README: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_dark_coco-wholebody.md
 Models:
 - Config: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_res50_coco_wholebody_256x192_dark.py
-  In Collection: wholebody--2d_kpt_sview_rgb_img--topdown_heatmap--coco-wholebody--vipnas_dark_coco-wholebody
-    [ViPNAS+DarkPose@COCO-WholeBody]
+  In Collection: ViPNAS
   Metadata:
+    Architecture:
+    - ViPNAS
+    - DarkPose
     Training Data: COCO-WholeBody
   Name: topdown_heatmap_vipnas_res50_coco_wholebody_256x192_dark
+  README: configs/wholebody/2d_kpt_sview_rgb_img/topdown_heatmap/coco-wholebody/vipnas_dark_coco-wholebody.md
   Results:
   - Dataset: COCO-WholeBody
     Metrics:

mmpose/models/backbones/__init__.py

@@ -20,11 +20,13 @@
 from .shufflenet_v2 import ShuffleNetV2
 from .tcn import TCN
 from .vgg import VGG
+from .vipnas_mbv3 import ViPNAS_MobileNetV3
 from .vipnas_resnet import ViPNAS_ResNet
 
 __all__ = [
     'AlexNet', 'HourglassNet', 'HourglassAENet', 'HRNet', 'MobileNetV2',
     'MobileNetV3', 'RegNet', 'ResNet', 'ResNetV1d', 'ResNeXt', 'SCNet',
     'SEResNet', 'SEResNeXt', 'ShuffleNetV1', 'ShuffleNetV2', 'CPM', 'RSN',
-    'MSPN', 'ResNeSt', 'VGG', 'TCN', 'ViPNAS_ResNet', 'LiteHRNet'
+    'MSPN', 'ResNeSt', 'VGG', 'TCN', 'ViPNAS_ResNet', 'ViPNAS_MobileNetV3',
+    'LiteHRNet'
 ]

mmpose/models/backbones/utils/inverted_residual.py

@@ -17,6 +17,8 @@ class InvertedResidual(nn.Module):
         mid_channels (int): The input channels of the depthwise convolution.
         kernel_size (int): The kernel size of the depthwise convolution.
             Default: 3.
+        groups (None or int): The group number of the depthwise convolution.
+            Default: None, which means group number = mid_channels.
         stride (int): The stride of the depthwise convolution. Default: 1.
         se_cfg (dict): Config dict for se layer. Default: None, which means no
             se layer.
@@ -41,6 +43,7 @@ def __init__(self,
                  out_channels,
                  mid_channels,
                  kernel_size=3,
+                 groups=None,
                  stride=1,
                  se_cfg=None,
                  with_expand_conv=True,
@@ -58,6 +61,9 @@ def __init__(self,
         self.with_se = se_cfg is not None
         self.with_expand_conv = with_expand_conv
 
+        if groups is None:
+            groups = mid_channels
+
         if self.with_se:
             assert isinstance(se_cfg, dict)
         if not self.with_expand_conv:
@@ -79,7 +85,7 @@ def __init__(self,
             kernel_size=kernel_size,
             stride=stride,
             padding=kernel_size // 2,
-            groups=mid_channels,
+            groups=groups,
             conv_cfg=conv_cfg,
             norm_cfg=norm_cfg,
             act_cfg=act_cfg)
@@ -93,7 +99,7 @@ def __init__(self,
             padding=0,
             conv_cfg=conv_cfg,
             norm_cfg=norm_cfg,
-            act_cfg=act_cfg)
+            act_cfg=None)
 
     def forward(self, x):