Merge branch 'sun_22.10' into 'main'

Sun 22.10

See merge request dcutoolkit/deeplearing/dlexamples_new!54

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_hmdb51_split1_keypoint.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(3, 4, 6),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=51,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file = 'data/posec3d/hmdb51.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(48, 48), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 56)),
+    dict(type='CenterCrop', crop_size=56),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 56)),
+    dict(type='CenterCrop', crop_size=56),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=10,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file,
+            split='train1',
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file,
+        split='test1',
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file,
+        split='test1',
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.01, momentum=0.9,
+    weight_decay=0.0001)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='step', step=[9, 11])
+total_epochs = 12
+checkpoint_config = dict(interval=1)
+workflow = [('train', 1)]
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'], topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d_iclr/slowonly_kinetics400_pretrained_r50_u48_120e_hmdb51_split1_keypoint'  # noqa: E501
+load_from = 'https://download.openmmlab.com/mmaction/skeleton/posec3d/k400_posec3d-041f49c6.pth'  # noqa: E501
+resume_from = None
+find_unused_parameters = True

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_ucf101_split1_keypoint.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(3, 4, 6),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=101,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file = 'data/posec3d/ucf101.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(48, 48), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 56)),
+    dict(type='CenterCrop', crop_size=56),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 56)),
+    dict(type='CenterCrop', crop_size=56),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=10,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file,
+            split='train1',
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file,
+        split='test1',
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file,
+        split='test1',
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.01, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='step', step=[9, 11])
+total_epochs = 12
+checkpoint_config = dict(interval=1)
+workflow = [('train', 1)]
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'], topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d_iclr/slowonly_kinetics400_pretrained_r50_u48_120e_ucf101_split1_keypoint'  # noqa: E501
+load_from = 'https://download.openmmlab.com/mmaction/skeleton/posec3d/k400_posec3d-041f49c6.pth'  # noqa: E501
+resume_from = None
+find_unused_parameters = True

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_gym_keypoint.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=99,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/gym_train.pkl'
+ann_file_val = 'data/posec3d/gym_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_gym_keypoint'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_gym_limb.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=99,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/gym_train.pkl'
+ann_file_val = 'data/posec3d/gym_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+skeletons = [[0, 5], [0, 6], [5, 7], [7, 9], [6, 8], [8, 10], [5, 11],
+             [11, 13], [13, 15], [6, 12], [12, 14], [14, 16], [0, 1], [0, 2],
+             [1, 3], [2, 4], [11, 12]]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_gym_limb'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=120,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/ntu120_xsub_train.pkl'
+ann_file_val = 'data/posec3d/ntu120_xsub_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        class_prob={i: 1 + int(i >= 60)
+                    for i in range(120)},
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_limb.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=120,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/ntu60_xsub_train.pkl'
+ann_file_val = 'data/posec3d/ntu60_xsub_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+skeletons = [[0, 5], [0, 6], [5, 7], [7, 9], [6, 8], [8, 10], [5, 11],
+             [11, 13], [13, 15], [6, 12], [12, 14], [14, 16], [0, 1], [0, 2],
+             [1, 3], [2, 4], [11, 12]]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        class_prob={i: 1 + int(i >= 60)
+                    for i in range(120)},
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_ntu120_xsub_limb'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=60,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/ntu60_xsub_train.pkl'
+ann_file_val = 'data/posec3d/ntu60_xsub_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=True,
+        with_limb=False,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu60_xsub_limb.py

+model = dict(
+    type='Recognizer3D',
+    backbone=dict(
+        type='ResNet3dSlowOnly',
+        depth=50,
+        pretrained=None,
+        in_channels=17,
+        base_channels=32,
+        num_stages=3,
+        out_indices=(2, ),
+        stage_blocks=(4, 6, 3),
+        conv1_stride_s=1,
+        pool1_stride_s=1,
+        inflate=(0, 1, 1),
+        spatial_strides=(2, 2, 2),
+        temporal_strides=(1, 1, 2),
+        dilations=(1, 1, 1)),
+    cls_head=dict(
+        type='I3DHead',
+        in_channels=512,
+        num_classes=60,
+        spatial_type='avg',
+        dropout_ratio=0.5),
+    train_cfg=dict(),
+    test_cfg=dict(average_clips='prob'))
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/ntu60_xsub_train.pkl'
+ann_file_val = 'data/posec3d/ntu60_xsub_val.pkl'
+left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
+right_kp = [2, 4, 6, 8, 10, 12, 14, 16]
+skeletons = [[0, 5], [0, 6], [5, 7], [7, 9], [6, 8], [8, 10], [5, 11],
+             [11, 13], [13, 15], [6, 12], [12, 14], [14, 16], [0, 1], [0, 2],
+             [1, 3], [2, 4], [11, 12]]
+train_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='RandomResizedCrop', area_range=(0.56, 1.0)),
+    dict(type='Resize', scale=(56, 56), keep_ratio=False),
+    dict(type='Flip', flip_ratio=0.5, left_kp=left_kp, right_kp=right_kp),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs', 'label'])
+]
+val_pipeline = [
+    dict(type='UniformSampleFrames', clip_len=48, num_clips=1, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+test_pipeline = [
+    dict(
+        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True),
+    dict(type='PoseDecode'),
+    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
+    dict(type='Resize', scale=(-1, 64)),
+    dict(type='CenterCrop', crop_size=64),
+    dict(
+        type='GeneratePoseTarget',
+        sigma=0.6,
+        use_score=True,
+        with_kp=False,
+        with_limb=True,
+        skeletons=skeletons,
+        double=True,
+        left_kp=left_kp,
+        right_kp=right_kp),
+    dict(type='FormatShape', input_format='NCTHW'),
+    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['imgs'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.2, momentum=0.9,
+    weight_decay=0.0003)  # this lr is used for 8 gpus
+optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
+# learning policy
+lr_config = dict(policy='CosineAnnealing', by_epoch=False, min_lr=0)
+total_epochs = 240
+checkpoint_config = dict(interval=10)
+workflow = [('train', 10)]
+evaluation = dict(
+    interval=10,
+    metrics=['top_k_accuracy', 'mean_class_accuracy'],
+    topk=(1, 5))
+log_config = dict(
+    interval=20, hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/posec3d/slowonly_r50_u48_240e_ntu60_xsub_limb'
+load_from = None
+resume_from = None
+find_unused_parameters = False

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/README.md

+# STGCN
+
+[Spatial temporal graph convolutional networks for skeleton-based action recognition](https://ojs.aaai.org/index.php/AAAI/article/view/12328)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+<!-- [ABSTRACT] -->
+
+Dynamics of human body skeletons convey significant information for human action recognition. Conventional approaches for modeling skeletons usually rely on hand-crafted parts or traversal rules, thus resulting in limited expressive power and difficulties of generalization. In this work, we propose a novel model of dynamic skeletons called Spatial-Temporal Graph Convolutional Networks (ST-GCN), which moves beyond the limitations of previous methods by automatically learning both the spatial and temporal patterns from data. This formulation not only leads to greater expressive power but also stronger generalization capability. On two large datasets, Kinetics and NTU-RGBD, it achieves substantial improvements over mainstream methods.
+
+<!-- [IMAGE] -->
+
+<div align=center>
+<img src="https://user-images.githubusercontent.com/34324155/142995893-d6618728-072c-46e1-b276-9b88cf21a01c.png" width="800"/>
+</div>
+
+## Results and Models
+
+### NTU60_XSub
+
+| config                                                                                          | keypoint | gpus | backbone | Top-1 |                                                                     ckpt                                                                      |                                                                 log                                                                 |                                                                 json                                                                  |
+| :---------------------------------------------------------------------------------------------- | :------: | :--: | :------: | :---: | :-------------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------: |
+| [stgcn_80e_ntu60_xsub_keypoint](/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py)       |    2d    |  2   |  STGCN   | 86.91 |    [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint-e7bb9653.pth)    |    [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.log)    |    [json](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.json)    |
+| [stgcn_80e_ntu60_xsub_keypoint_3d](/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d.py) |    3d    |  1   |  STGCN   | 84.61 | [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d-13e7ccf0.pth) | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.log) | [json](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.json) |
+
+### BABEL
+
+| config                                                                      | gpus | backbone |   Top-1   | Mean Top-1 | Top-1 Official (AGCN) | Mean Top-1 Official (AGCN) |                                                           ckpt                                                            |                                                    log                                                     |
+| --------------------------------------------------------------------------- | :--: | :------: | :-------: | :--------: | :-------------------: | :------------------------: | :-----------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------: |
+| [stgcn_80e_babel60](/configs/skeleton/stgcn/stgcn_80e_babel60.py)           |  8   |  ST-GCN  | **42.39** | **28.28**  |         41.14         |           24.46            |      [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60-3d206418.pth)      |   [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60.log)    |
+| [stgcn_80e_babel60_wfl](/configs/skeleton/stgcn/stgcn_80e_babel60_wfl.py)   |  8   |  ST-GCN  | **40.31** |   29.79    |         33.41         |         **30.42**          |  [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60_wfl/stgcn_80e_babel60_wfl-1a9102d7.pth)  | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60_wfl.log)  |
+| [stgcn_80e_babel120](/configs/skeleton/stgcn/stgcn_80e_babel120.py)         |  8   |  ST-GCN  | **38.95** | **20.58**  |         38.41         |           17.56            |     [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120/stgcn_80e_babel120-e41eb6d7.pth)     |   [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel120.log)   |
+| [stgcn_80e_babel120_wfl](/configs/skeleton/stgcn/stgcn_80e_babel120_wfl.py) |  8   |  ST-GCN  | **33.00** |   24.33    |         27.91         |        **26.17**\*         | [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120_wfl/stgcn_80e_babel120_wfl-3f2c100d.pth) | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel120_wfl.log) |
+
+\* The number is copied from the [paper](https://arxiv.org/pdf/2106.09696.pdf), the performance of the [released checkpoints](https://github.com/abhinanda-punnakkal/BABEL/tree/main/action_recognition) for BABEL-120 is inferior.
+
+## Train
+
+You can use the following command to train a model.
+
+```shell
+python tools/train.py ${CONFIG_FILE} [optional arguments]
+```
+
+Example: train STGCN model on NTU60 dataset in a deterministic option with periodic validation.
+
+```shell
+python tools/train.py configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py \
+    --work-dir work_dirs/stgcn_80e_ntu60_xsub_keypoint \
+    --validate --seed 0 --deterministic
+```
+
+For more details, you can refer to **Training setting** part in [getting_started](/docs/getting_started.md#training-setting).
+
+## Test
+
+You can use the following command to test a model.
+
+```shell
+python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]
+```
+
+Example: test STGCN model on NTU60 dataset and dump the result to a pickle file.
+
+```shell
+python tools/test.py configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py \
+    checkpoints/SOME_CHECKPOINT.pth --eval top_k_accuracy mean_class_accuracy \
+    --out result.pkl
+```
+
+For more details, you can refer to **Test a dataset** part in [getting_started](/docs/getting_started.md#test-a-dataset).
+
+## Citation
+
+```BibTeX
+@inproceedings{yan2018spatial,
+  title={Spatial temporal graph convolutional networks for skeleton-based action recognition},
+  author={Yan, Sijie and Xiong, Yuanjun and Lin, Dahua},
+  booktitle={Thirty-second AAAI conference on artificial intelligence},
+  year={2018}
+}
+```

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/README_zh-CN.md

+# STGCN
+
+## 简介
+
+<!-- [ALGORITHM] -->
+
+```BibTeX
+@inproceedings{yan2018spatial,
+  title={Spatial temporal graph convolutional networks for skeleton-based action recognition},
+  author={Yan, Sijie and Xiong, Yuanjun and Lin, Dahua},
+  booktitle={Thirty-second AAAI conference on artificial intelligence},
+  year={2018}
+}
+```
+
+## 模型库
+
+### NTU60_XSub
+
+| 配置文件                                                                                        | 骨骼点 | GPU 数量 | 主干网络 | Top-1 准确率 |                                                                     ckpt                                                                      |                                                                 log                                                                 |                                                                 json                                                                  |
+| :---------------------------------------------------------------------------------------------- | :----: | :------: | :------: | :----------: | :-------------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------: |
+| [stgcn_80e_ntu60_xsub_keypoint](/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py)       |   2d   |    2     |  STGCN   |    86.91     |    [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint-e7bb9653.pth)    |    [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.log)    |    [json](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.json)    |
+| [stgcn_80e_ntu60_xsub_keypoint_3d](/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d.py) |   3d   |    1     |  STGCN   |    84.61     | [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d-13e7ccf0.pth) | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.log) | [json](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.json) |
+
+### BABEL
+
+| 配置文件                                                                    | GPU 数量 | 主干网络 | Top-1 准确率 | 类平均 Top-1 准确率 | Top-1 准确率 <br>（官方，使用 AGCN） | 类平均 Top-1 准确率<br>（官方，使用 AGCN） |                                                           ckpt                                                            |                                                    log                                                     |
+| --------------------------------------------------------------------------- | :------: | :------: | :----------: | :-----------------: | :----------------------------------: | :----------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------: |
+| [stgcn_80e_babel60](/configs/skeleton/stgcn/stgcn_80e_babel60.py)           |    8     |  ST-GCN  |  **42.39**   |      **28.28**      |                41.14                 |                   24.46                    |      [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60-3d206418.pth)      |   [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60.log)    |
+| [stgcn_80e_babel60_wfl](/configs/skeleton/stgcn/stgcn_80e_babel60_wfl.py)   |    8     |  ST-GCN  |  **40.31**   |        29.79        |                33.41                 |                 **30.42**                  |  [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60_wfl/stgcn_80e_babel60_wfl-1a9102d7.pth)  | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60_wfl.log)  |
+| [stgcn_80e_babel120](/configs/skeleton/stgcn/stgcn_80e_babel120.py)         |    8     |  ST-GCN  |  **38.95**   |      **20.58**      |                38.41                 |                   17.56                    |     [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120/stgcn_80e_babel120-e41eb6d7.pth)     |   [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel120.log)   |
+| [stgcn_80e_babel120_wfl](/configs/skeleton/stgcn/stgcn_80e_babel120_wfl.py) |    8     |  ST-GCN  |  **33.00**   |        24.33        |                27.91                 |                **26.17**\*                 | [ckpt](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120_wfl/stgcn_80e_babel120_wfl-3f2c100d.pth) | [log](https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel120_wfl.log) |
+
+\* 注：此数字引自原 [论文](https://arxiv.org/pdf/2106.09696.pdf)， 实际公开的 [模型权重](https://github.com/abhinanda-punnakkal/BABEL/tree/main/action_recognition) 精度略低一些。
+
+## 如何训练
+
+用户可以使用以下指令进行模型训练。
+
+```shell
+python tools/train.py ${CONFIG_FILE} [optional arguments]
+```
+
+例如：以一个确定性的训练方式，辅以定期的验证过程进行 STGCN 模型在 NTU60 数据集上的训练
+
+```shell
+python tools/train.py configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py \
+    --work-dir work_dirs/stgcn_80e_ntu60_xsub_keypoint \
+    --validate --seed 0 --deterministic
+```
+
+更多训练细节，可参考 [基础教程](/docs_zh_CN/getting_started.md#%E8%AE%AD%E7%BB%83%E9%85%8D%E7%BD%AE) 中的 **训练配置** 部分。
+
+## 如何测试
+
+用户可以使用以下指令进行模型测试。
+
+```shell
+python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]
+```
+
+例如：在 NTU60 数据集上测试 STGCN 模型，并将结果导出为一个 pickle 文件。
+
+```shell
+python tools/test.py configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py \
+    checkpoints/SOME_CHECKPOINT.pth --eval top_k_accuracy mean_class_accuracy \
+    --out result.pkl
+```
+
+更多测试细节，可参考 [基础教程](/docs_zh_CN/getting_started.md#%E6%B5%8B%E8%AF%95%E6%9F%90%E4%B8%AA%E6%95%B0%E6%8D%AE%E9%9B%86) 中的 **测试某个数据集** 部分。

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/metafile.yml

+Collections:
+- Name: STGCN
+  README: configs/skeleton/stgcn/README.md
+Models:
+- Config: configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 16
+    Epochs: 80
+    Parameters: 3088704
+    Training Data: NTU60-XSub
+    Training Resources: 2 GPUs
+  Name: stgcn_80e_ntu60_xsub_keypoint
+  Results:
+    Dataset: NTU60-XSub
+    Metrics:
+      Top 1 Accuracy: 86.91
+    Task: Skeleton-based Action Recognition
+  Training Json Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.json
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint-e7bb9653.pth
+- Config: configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 32
+    Epochs: 80
+    Parameters: 3088704
+    Training Data: NTU60-XSub
+    Training Resources: 1 GPU
+  Name: stgcn_80e_ntu60_xsub_keypoint_3d
+  Results:
+    Dataset: NTU60-XSub
+    Metrics:
+      Top 1 Accuracy: 84.61
+    Task: Skeleton-based Action Recognition
+  Training Json Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.json
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d/stgcn_80e_ntu60_xsub_keypoint_3d-13e7ccf0.pth
+- Config: configs/skeleton/stgcn/stgcn_80e_babel60.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 128
+    Epochs: 80
+    Parameters: 3088704
+    Training Data: BABEL60
+    Training Resources: 8 GPU
+  Name: stgcn_80e_babel60
+  Results:
+    Dataset: BABEL60
+    Metrics:
+      Top 1 Accuracy: 42.39
+      Mean Top 1 Accuracy: 28.28
+    Task: Skeleton-based Action Recognition
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60/stgcn_80e_babel60-3d206418.pth
+- Config: configs/skeleton/stgcn/stgcn_80e_babel60_wfl.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 128
+    Epochs: 80
+    Parameters: 3088704
+    Training Data: BABEL60
+    Training Resources: 8 GPU
+  Name: stgcn_80e_babel60_wfl
+  Results:
+    Dataset: BABEL60
+    Metrics:
+      Top 1 Accuracy: 40.31
+      Mean Top 1 Accuracy: 29.79
+    Task: Skeleton-based Action Recognition
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60_wfl/stgcn_80e_babel60_wfl.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel60_wfl/stgcn_80e_babel60_wfl-1a9102d7.pth
+- Config: configs/skeleton/stgcn/stgcn_80e_babel120.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 128
+    Epochs: 80
+    Parameters: 3104320
+    Training Data: BABEL120
+    Training Resources: 8 GPU
+  Name: stgcn_80e_babel120
+  Results:
+    Dataset: BABEL120
+    Metrics:
+      Top 1 Accuracy: 38.95
+      Mean Top 1 Accuracy: 20.58
+    Task: Skeleton-based Action Recognition
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120/stgcn_80e_babel120.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120/stgcn_80e_babel120-e41eb6d7.pth
+- Config: configs/skeleton/stgcn/stgcn_80e_babel120_wfl.py
+  In Collection: STGCN
+  Metadata:
+    Architecture: STGCN
+    Batch Size: 128
+    Epochs: 80
+    Parameters: 3104320
+    Training Data: BABEL120
+    Training Resources: 8 GPU
+  Name: stgcn_80e_babel120_wfl
+  Results:
+    Dataset: BABEL120
+    Metrics:
+      Top 1 Accuracy: 33.00
+      Mean Top 1 Accuracy: 24.33
+    Task: Skeleton-based Action Recognition
+  Training Log: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120_wfl/stgcn_80e_babel120_wfl.log
+  Weights: https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_babel120_wfl/stgcn_80e_babel120_wfl-3f2c100d.pth

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_babel120.py

+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='ntu-rgb+d', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=120,
+        in_channels=256,
+        num_person=1,
+        loss_cls=dict(type='CrossEntropyLoss')),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/babel/babel120_train.pkl'
+ann_file_val = 'data/babel/babel120_val.pkl'
+train_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=5,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file_train,
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 14])
+total_epochs = 16
+checkpoint_config = dict(interval=1)
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_babel120'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_babel120_wfl.py

+samples_per_cls = [
+    518, 1993, 6260, 508, 208, 3006, 431, 724, 4527, 2131, 199, 1255, 487, 302,
+    136, 571, 267, 646, 1180, 405, 72, 731, 842, 1619, 271, 27, 1198, 1012,
+    110, 865, 462, 526, 405, 487, 101, 24, 84, 64, 168, 271, 609, 503, 76, 167,
+    415, 137, 421, 283, 2069, 715, 196, 66, 44, 989, 122, 43, 599, 396, 245,
+    380, 34, 236, 260, 325, 127, 133, 119, 66, 125, 50, 206, 191, 394, 69, 98,
+    145, 38, 21, 29, 64, 277, 65, 39, 31, 35, 85, 54, 80, 133, 66, 39, 64, 268,
+    34, 172, 54, 33, 21, 110, 19, 40, 55, 146, 39, 37, 75, 101, 20, 46, 55, 43,
+    21, 43, 87, 29, 36, 24, 37, 28, 39
+]
+
+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='ntu-rgb+d', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=120,
+        in_channels=256,
+        num_person=1,
+        loss_cls=dict(type='CBFocalLoss', samples_per_cls=samples_per_cls)),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/babel/babel120_train.pkl'
+ann_file_val = 'data/babel/babel120_val.pkl'
+train_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=5,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file_train,
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 14])
+total_epochs = 16
+checkpoint_config = dict(interval=1)
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_babel120_wfl/'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_babel60.py

+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='ntu-rgb+d', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=60,
+        in_channels=256,
+        num_person=1,
+        loss_cls=dict(type='CrossEntropyLoss')),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/babel/babel60_train.pkl'
+ann_file_val = 'data/babel/babel60_val.pkl'
+train_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=5,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file_train,
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 14])
+total_epochs = 16
+checkpoint_config = dict(interval=1)
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_babel60'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_babel60_wfl.py

+samples_per_cls = [
+    518, 1993, 6260, 508, 208, 3006, 431, 724, 4527, 2131, 199, 1255, 487, 302,
+    136, 571, 267, 646, 1180, 405, 731, 842, 1619, 271, 1198, 1012, 865, 462,
+    526, 405, 487, 168, 271, 609, 503, 167, 415, 421, 283, 2069, 715, 196, 989,
+    122, 599, 396, 245, 380, 236, 260, 325, 133, 206, 191, 394, 145, 277, 268,
+    172, 146
+]
+
+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='ntu-rgb+d', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=60,
+        in_channels=256,
+        num_person=1,
+        loss_cls=dict(type='CBFocalLoss', samples_per_cls=samples_per_cls)),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/babel/babel60_train.pkl'
+ann_file_val = 'data/babel/babel60_val.pkl'
+train_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM', num_person=1),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type='RepeatDataset',
+        times=5,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=ann_file_train,
+            data_prefix='',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 14])
+total_epochs = 16
+checkpoint_config = dict(interval=1)
+evaluation = dict(
+    interval=1, metrics=['top_k_accuracy', 'mean_class_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_babel60_wfl/'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py

+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='coco', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=60,
+        in_channels=256,
+        loss_cls=dict(type='CrossEntropyLoss')),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/posec3d/ntu60_xsub_train.pkl'
+ann_file_val = 'data/posec3d/ntu60_xsub_val.pkl'
+train_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='PoseNormalize'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='PoseNormalize'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='PoseNormalize'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=16,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 50])
+total_epochs = 80
+checkpoint_config = dict(interval=5)
+evaluation = dict(interval=5, metrics=['top_k_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_ntu60_xsub_keypoint/'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint_3d.py

+model = dict(
+    type='SkeletonGCN',
+    backbone=dict(
+        type='STGCN',
+        in_channels=3,
+        edge_importance_weighting=True,
+        graph_cfg=dict(layout='ntu-rgb+d', strategy='spatial')),
+    cls_head=dict(
+        type='STGCNHead',
+        num_classes=60,
+        in_channels=256,
+        loss_cls=dict(type='CrossEntropyLoss')),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'PoseDataset'
+ann_file_train = 'data/ntu/nturgb+d_skeletons_60_3d_nmtvc/xsub/train.pkl'
+ann_file_val = 'data/ntu/nturgb+d_skeletons_60_3d_nmtvc/xsub/val.pkl'
+train_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+val_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+test_pipeline = [
+    dict(type='PaddingWithLoop', clip_len=300),
+    dict(type='PoseDecode'),
+    dict(type='FormatGCNInput', input_format='NCTVM'),
+    dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
+    dict(type='ToTensor', keys=['keypoint'])
+]
+data = dict(
+    videos_per_gpu=32,
+    workers_per_gpu=2,
+    test_dataloader=dict(videos_per_gpu=1),
+    train=dict(
+        type=dataset_type,
+        ann_file=ann_file_train,
+        data_prefix='',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=val_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=ann_file_val,
+        data_prefix='',
+        pipeline=test_pipeline))
+
+# optimizer
+optimizer = dict(
+    type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001, nesterov=True)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(policy='step', step=[10, 50])
+total_epochs = 80
+checkpoint_config = dict(interval=3)
+evaluation = dict(interval=3, metrics=['top_k_accuracy'])
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+
+# runtime settings
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+work_dir = './work_dirs/stgcn_80e_ntu60_xsub_keypoint_3d/'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]

openmmlab_test/mmaction2-0.24.1/demo/README.md

+# Demo
+
+## Outline
+
+- [Modify configs through script arguments](#modify-config-through-script-arguments): Tricks to directly modify configs through script arguments.
+- [Video demo](#video-demo): A demo script to predict the recognition result using a single video.
+- [SpatioTemporal Action Detection Video Demo](#spatiotemporal-action-detection-video-demo): A demo script to predict the SpatioTemporal Action Detection result using a single video.
+- [Video GradCAM Demo](#video-gradcam-demo): A demo script to visualize GradCAM results using a single video.
+- [Webcam demo](#webcam-demo): A demo script to implement real-time action recognition from a web camera.
+- [Long Video demo](#long-video-demo): a demo script to predict different labels using a single long video.
+- [SpatioTemporal Action Detection Webcam Demo](#spatiotemporal-action-detection-webcam-demo): A demo script to implement real-time spatio-temporal action detection from a web camera.
+- [Skeleton-based Action Recognition Demo](#skeleton-based-action-recognition-demo): A demo script to predict the skeleton-based action recognition result using a single video.
+- [Video Structuralize Demo](#video-structuralize-demo): A demo script to predict the skeleton-based and rgb-based action recognition and spatio-temporal action detection result using a single video.
+- [Audio Demo](#audio-demo): A demo script to predict the recognition result using a single audio file.
+
+## Modify configs through script arguments
+
+When running demos using our provided scripts, you may specify `--cfg-options` to in-place modify the config.
+
+- Update config keys of dict.
+
+  The config options can be specified following the order of the dict keys in the original config.
+  For example, `--cfg-options model.backbone.norm_eval=False` changes the all BN modules in model backbones to `train` mode.
+
+- Update keys inside a list of configs.
+
+  Some config dicts are composed as a list in your config. For example, the training pipeline `data.train.pipeline` is normally a list
+  e.g. `[dict(type='SampleFrames'), ...]`. If you want to change `'SampleFrames'` to `'DenseSampleFrames'` in the pipeline,
+  you may specify `--cfg-options data.train.pipeline.0.type=DenseSampleFrames`.
+
+- Update values of list/tuples.
+
+  If the value to be updated is a list or a tuple. For example, the config file normally sets `workflow=[('train', 1)]`. If you want to
+  change this key, you may specify `--cfg-options workflow="[(train,1),(val,1)]"`. Note that the quotation mark " is necessary to
+  support list/tuple data types, and that **NO** white space is allowed inside the quotation marks in the specified value.
+
+## Video demo
+
+We provide a demo script to predict the recognition result using a single video. In order to get predict results in range `[0, 1]`, make sure to set `model['test_cfg'] = dict(average_clips='prob')` in config file.
+
+```shell
+python demo/demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${VIDEO_FILE} {LABEL_FILE} [--use-frames] \
+    [--device ${DEVICE_TYPE}] [--fps {FPS}] [--font-scale {FONT_SCALE}] [--font-color {FONT_COLOR}] \
+    [--target-resolution ${TARGET_RESOLUTION}] [--resize-algorithm {RESIZE_ALGORITHM}] [--out-filename {OUT_FILE}]
+```
+
+Optional arguments:
+
+- `--use-frames`: If specified, the demo will take rawframes as input. Otherwise, it will take a video as input.
+- `DEVICE_TYPE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`. If not specified, it will be set to `cuda:0`.
+- `FPS`: FPS value of the output video when using rawframes as input. If not specified, it will be set to 30.
+- `FONT_SCALE`: Font scale of the label added in the video. If not specified, it will be 0.5.
+- `FONT_COLOR`: Font color of the label added in the video. If not specified, it will be `white`.
+- `TARGET_RESOLUTION`: Resolution(desired_width, desired_height) for resizing the frames before output when using a video as input. If not specified, it will be None and the frames are resized by keeping the existing aspect ratio.
+- `RESIZE_ALGORITHM`: Resize algorithm used for resizing. If not specified, it will be set to `bicubic`.
+- `OUT_FILE`: Path to the output file which can be a video format or gif format. If not specified, it will be set to `None` and does not generate the output file.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` and have already downloaded the checkpoints to the directory `checkpoints/`,
+or use checkpoint url from `configs/` to directly load corresponding checkpoint, which will be automatically saved in `$HOME/.cache/torch/checkpoints`.
+
+1. Recognize a video file as input by using a TSN model on cuda by default.
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt
+   ```
+
+2. Recognize a video file as input by using a TSN model on cuda by default, loading checkpoint from url.
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       https://download.openmmlab.com/mmaction/recognition/tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt
+   ```
+
+3. Recognize a list of rawframes as input by using a TSN model on cpu.
+
+   ```shell
+   python demo/demo.py configs/recognition/tsn/tsn_r50_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       PATH_TO_FRAMES/ LABEL_FILE --use-frames --device cpu
+   ```
+
+4. Recognize a video file as input by using a TSN model and then generate an mp4 file.
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt --out-filename demo/demo_out.mp4
+   ```
+
+5. Recognize a list of rawframes as input by using a TSN model and then generate a gif file.
+
+   ```shell
+   python demo/demo.py configs/recognition/tsn/tsn_r50_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       PATH_TO_FRAMES/ LABEL_FILE --use-frames --out-filename demo/demo_out.gif
+   ```
+
+6. Recognize a video file as input by using a TSN model, then generate an mp4 file with a given resolution and resize algorithm.
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt --target-resolution 340 256 --resize-algorithm bilinear \
+       --out-filename demo/demo_out.mp4
+   ```
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   # If either dimension is set to -1, the frames are resized by keeping the existing aspect ratio
+   # For --target-resolution 170 -1, original resolution (340, 256) -> target resolution (170, 128)
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt --target-resolution 170 -1 --resize-algorithm bilinear \
+       --out-filename demo/demo_out.mp4
+   ```
+
+7. Recognize a video file as input by using a TSN model, then generate an mp4 file with a label in a red color and fontscale 1.
+
+   ```shell
+   # The demo.mp4 and label_map_k400.txt are both from Kinetics-400
+   python demo/demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       demo/demo.mp4 tools/data/kinetics/label_map_k400.txt --font-scale 1 --font-color red \
+       --out-filename demo/demo_out.mp4
+   ```
+
+8. Recognize a list of rawframes as input by using a TSN model and then generate an mp4 file with 24 fps.
+
+   ```shell
+   python demo/demo.py configs/recognition/tsn/tsn_r50_inference_1x1x3_100e_kinetics400_rgb.py \
+       checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+       PATH_TO_FRAMES/ LABEL_FILE --use-frames --fps 24 --out-filename demo/demo_out.gif
+   ```
+
+## SpatioTemporal Action Detection Video Demo
+
+We provide a demo script to predict the SpatioTemporal Action Detection result using a single video.
+
+```shell
+python demo/demo_spatiotemporal_det.py --video ${VIDEO_FILE} \
+    [--config ${SPATIOTEMPORAL_ACTION_DETECTION_CONFIG_FILE}] \
+    [--checkpoint ${SPATIOTEMPORAL_ACTION_DETECTION_CHECKPOINT}] \
+    [--det-config ${HUMAN_DETECTION_CONFIG_FILE}] \
+    [--det-checkpoint ${HUMAN_DETECTION_CHECKPOINT}] \
+    [--det-score-thr ${HUMAN_DETECTION_SCORE_THRESHOLD}] \
+    [--action-score-thr ${ACTION_DETECTION_SCORE_THRESHOLD}] \
+    [--label-map ${LABEL_MAP}] \
+    [--device ${DEVICE}] \
+    [--out-filename ${OUTPUT_FILENAME}] \
+    [--predict-stepsize ${PREDICT_STEPSIZE}] \
+    [--output-stepsize ${OUTPUT_STEPSIZE}] \
+    [--output-fps ${OUTPUT_FPS}]
+```
+
+Optional arguments:
+
+- `SPATIOTEMPORAL_ACTION_DETECTION_CONFIG_FILE`: The spatiotemporal action detection config file path.
+- `SPATIOTEMPORAL_ACTION_DETECTION_CHECKPOINT`: The spatiotemporal action detection checkpoint URL.
+- `HUMAN_DETECTION_CONFIG_FILE`: The human detection config file path.
+- `HUMAN_DETECTION_CHECKPOINT`: The human detection checkpoint URL.
+- `HUMAN_DETECTION_SCORE_THRE`: The score threshold for human detection. Default: 0.9.
+- `ACTION_DETECTION_SCORE_THRESHOLD`: The score threshold for action detection. Default: 0.5.
+- `LABEL_MAP`: The label map used. Default: `tools/data/ava/label_map.txt`.
+- `DEVICE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`.  Default: `cuda:0`.
+- `OUTPUT_FILENAME`: Path to the output file which is a video format. Default: `demo/stdet_demo.mp4`.
+- `PREDICT_STEPSIZE`: Make a prediction per N frames.  Default: 8.
+- `OUTPUT_STEPSIZE`: Output 1 frame per N frames in the input video. Note that `PREDICT_STEPSIZE % OUTPUT_STEPSIZE == 0`. Default: 4.
+- `OUTPUT_FPS`: The FPS of demo video output. Default: 6.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` .
+
+1. Use the Faster RCNN as the human detector, SlowOnly-8x8-R101 as the action detector. Making predictions per 8 frames, and output 1 frame per 4 frames to the output video. The FPS of the output video is 4.
+
+```shell
+python demo/demo_spatiotemporal_det.py --video demo/demo.mp4 \
+    --config configs/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb.py \
+    --checkpoint https://download.openmmlab.com/mmaction/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb_20201217-16378594.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --det-score-thr 0.9 \
+    --action-score-thr 0.5 \
+    --label-map tools/data/ava/label_map.txt \
+    --predict-stepsize 8 \
+    --output-stepsize 4 \
+    --output-fps 6
+```
+
+## Video GradCAM Demo
+
+We provide a demo script to visualize GradCAM results using a single video.
+
+```shell
+python demo/demo_gradcam.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${VIDEO_FILE} [--use-frames] \
+    [--device ${DEVICE_TYPE}] [--target-layer-name ${TARGET_LAYER_NAME}] [--fps {FPS}] \
+    [--target-resolution ${TARGET_RESOLUTION}] [--resize-algorithm {RESIZE_ALGORITHM}] [--out-filename {OUT_FILE}]
+```
+
+- `--use-frames`: If specified, the demo will take rawframes as input. Otherwise, it will take a video as input.
+- `DEVICE_TYPE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`. If not specified, it will be set to `cuda:0`.
+- `FPS`: FPS value of the output video when using rawframes as input. If not specified, it will be set to 30.
+- `OUT_FILE`: Path to the output file which can be a video format or gif format. If not specified, it will be set to `None` and does not generate the output file.
+- `TARGET_LAYER_NAME`: Layer name to generate GradCAM localization map.
+- `TARGET_RESOLUTION`: Resolution(desired_width, desired_height) for resizing the frames before output when using a video as input. If not specified, it will be None and the frames are resized by keeping the existing aspect ratio.
+- `RESIZE_ALGORITHM`: Resize algorithm used for resizing. If not specified, it will be set to `bilinear`.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` and have already downloaded the checkpoints to the directory `checkpoints/`,
+or use checkpoint url from `configs/` to directly load corresponding checkpoint, which will be automatically saved in `$HOME/.cache/torch/checkpoints`.
+
+1. Get GradCAM results of a I3D model, using a video file as input and then generate an gif file with 10 fps.
+
+   ```shell
+   python demo/demo_gradcam.py configs/recognition/i3d/i3d_r50_video_inference_32x2x1_100e_kinetics400_rgb.py \
+       checkpoints/i3d_r50_video_32x2x1_100e_kinetics400_rgb_20200826-e31c6f52.pth demo/demo.mp4 \
+       --target-layer-name backbone/layer4/1/relu --fps 10 \
+       --out-filename demo/demo_gradcam.gif
+   ```
+
+2. Get GradCAM results of a TSM model, using a video file as input and then generate an gif file, loading checkpoint from url.
+
+   ```shell
+   python demo/demo_gradcam.py configs/recognition/tsm/tsm_r50_video_inference_1x1x8_100e_kinetics400_rgb.py \
+       https://download.openmmlab.com/mmaction/recognition/tsm/tsm_r50_video_1x1x8_100e_kinetics400_rgb/tsm_r50_video_1x1x8_100e_kinetics400_rgb_20200702-a77f4328.pth \
+       demo/demo.mp4 --target-layer-name backbone/layer4/1/relu --out-filename demo/demo_gradcam_tsm.gif
+   ```
+
+## Webcam demo
+
+We provide a demo script to implement real-time action recognition from web camera. In order to get predict results in range `[0, 1]`, make sure to set `model.['test_cfg'] = dict(average_clips='prob')` in config file.
+
+```shell
+python demo/webcam_demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${LABEL_FILE} \
+    [--device ${DEVICE_TYPE}] [--camera-id ${CAMERA_ID}] [--threshold ${THRESHOLD}] \
+    [--average-size ${AVERAGE_SIZE}] [--drawing-fps ${DRAWING_FPS}] [--inference-fps ${INFERENCE_FPS}]
+```
+
+Optional arguments:
+
+- `DEVICE_TYPE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`. If not specified, it will be set to `cuda:0`.
+- `CAMERA_ID`: ID of camera device If not specified, it will be set to 0.
+- `THRESHOLD`: Threshold of prediction score for action recognition. Only label with score higher than the threshold will be shown. If not specified, it will be set to 0.
+- `AVERAGE_SIZE`: Number of latest clips to be averaged for prediction. If not specified, it will be set to 1.
+- `DRAWING_FPS`: Upper bound FPS value of the output drawing. If not specified, it will be set to 20.
+- `INFERENCE_FPS`: Upper bound FPS value of the output drawing. If not specified, it will be set to 4.
+
+:::{note}
+If your hardware is good enough, increasing the value of `DRAWING_FPS` and `INFERENCE_FPS` will get a better experience.
+:::
+
+Examples:
+
+Assume that you are located at `$MMACTION2` and have already downloaded the checkpoints to the directory `checkpoints/`,
+or use checkpoint url from `configs/` to directly load corresponding checkpoint, which will be automatically saved in `$HOME/.cache/torch/checkpoints`.
+
+1. Recognize the action from web camera as input by using a TSN model on cpu, averaging the score per 5 times
+   and outputting result labels with score higher than 0.2.
+
+   ```shell
+   python demo/webcam_demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+     checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth tools/data/kinetics/label_map_k400.txt --average-size 5 \
+     --threshold 0.2 --device cpu
+   ```
+
+2. Recognize the action from web camera as input by using a TSN model on cpu, averaging the score per 5 times
+   and outputting result labels with score higher than 0.2, loading checkpoint from url.
+
+   ```shell
+   python demo/webcam_demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+     https://download.openmmlab.com/mmaction/recognition/tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+     tools/data/kinetics/label_map_k400.txt --average-size 5 --threshold 0.2 --device cpu
+   ```
+
+3. Recognize the action from web camera as input by using a I3D model on gpu by default, averaging the score per 5 times
+   and outputting result labels with score higher than 0.2.
+
+   ```shell
+   python demo/webcam_demo.py configs/recognition/i3d/i3d_r50_video_inference_32x2x1_100e_kinetics400_rgb.py \
+     checkpoints/i3d_r50_32x2x1_100e_kinetics400_rgb_20200614-c25ef9a4.pth tools/data/kinetics/label_map_k400.txt \
+     --average-size 5 --threshold 0.2
+   ```
+
+:::{note}
+Considering the efficiency difference for users' hardware, Some modifications might be done to suit the case.
+Users can change:
+
+1). `SampleFrames` step (especially the number of `clip_len` and `num_clips`) of `test_pipeline` in the config file, like `--cfg-options data.test.pipeline.0.num_clips=3`.
+2). Change to the suitable Crop methods like `TenCrop`, `ThreeCrop`, `CenterCrop`, etc. in `test_pipeline` of the config file, like `--cfg-options data.test.pipeline.4.type=CenterCrop`.
+3). Change the number of `--average-size`. The smaller, the faster.
+:::
+
+## Long video demo
+
+We provide a demo script to predict different labels using a single long video. In order to get predict results in range `[0, 1]`, make sure to set `test_cfg = dict(average_clips='prob')` in config file.
+
+```shell
+python demo/long_video_demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${VIDEO_FILE} ${LABEL_FILE} \
+    ${OUT_FILE} [--input-step ${INPUT_STEP}] [--device ${DEVICE_TYPE}] [--threshold ${THRESHOLD}]
+```
+
+Optional arguments:
+
+- `OUT_FILE`: Path to the output, either video or json file
+- `INPUT_STEP`: Input step for sampling frames, which can help to get more spare input. If not specified , it will be set to 1.
+- `DEVICE_TYPE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`. If not specified, it will be set to `cuda:0`.
+- `THRESHOLD`: Threshold of prediction score for action recognition. Only label with score higher than the threshold will be shown. If not specified, it will be set to 0.01.
+- `STRIDE`: By default, the demo generates a prediction for each single frame, which might cost lots of time. To speed up, you can set the argument `STRIDE` and then the demo will generate a prediction every `STRIDE x sample_length` frames (`sample_length` indicates the size of temporal window from which you sample frames, which equals to `clip_len x frame_interval`). For example, if the sample_length is 64 frames and you set `STRIDE` to 0.5, predictions will be generated every 32 frames. If set as 0, predictions will be generated for each frame. The desired value of `STRIDE` is (0, 1\], while it also works for `STRIDE > 1` (the generated predictions will be too sparse). Default: 0.
+- `LABEL_COLOR`: Font Color of the labels in (B, G, R). Default is white, that is (256, 256, 256).
+- `MSG_COLOR`: Font Color of the messages in (B, G, R). Default is gray, that is (128, 128, 128).
+
+Examples:
+
+Assume that you are located at `$MMACTION2` and have already downloaded the checkpoints to the directory `checkpoints/`,
+or use checkpoint url from `configs/` to directly load corresponding checkpoint, which will be automatically saved in `$HOME/.cache/torch/checkpoints`.
+
+1. Predict different labels in a long video by using a TSN model on cpu, with 3 frames for input steps (that is, random sample one from each 3 frames)
+   and outputting result labels with score higher than 0.2.
+
+   ```shell
+   python demo/long_video_demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+     checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth PATH_TO_LONG_VIDEO tools/data/kinetics/label_map_k400.txt PATH_TO_SAVED_VIDEO \
+     --input-step 3 --device cpu --threshold 0.2
+   ```
+
+2. Predict different labels in a long video by using a TSN model on cpu, with 3 frames for input steps (that is, random sample one from each 3 frames)
+   and outputting result labels with score higher than 0.2, loading checkpoint from url.
+
+   ```shell
+   python demo/long_video_demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+     https://download.openmmlab.com/mmaction/recognition/tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+     PATH_TO_LONG_VIDEO tools/data/kinetics/label_map_k400.txt PATH_TO_SAVED_VIDEO --input-step 3 --device cpu --threshold 0.2
+   ```
+
+3. Predict different labels in a long video from web by using a TSN model on cpu, with 3 frames for input steps (that is, random sample one from each 3 frames)
+   and outputting result labels with score higher than 0.2, loading checkpoint from url.
+
+   ```shell
+   python demo/long_video_demo.py configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+     https://download.openmmlab.com/mmaction/recognition/tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+     https://www.learningcontainer.com/wp-content/uploads/2020/05/sample-mp4-file.mp4 \
+     tools/data/kinetics/label_map_k400.txt PATH_TO_SAVED_VIDEO --input-step 3 --device cpu --threshold 0.2
+   ```
+
+4. Predict different labels in a long video by using a I3D model on gpu, with input_step=1, threshold=0.01 as default and print the labels in cyan.
+
+   ```shell
+   python demo/long_video_demo.py configs/recognition/i3d/i3d_r50_video_inference_32x2x1_100e_kinetics400_rgb.py \
+     checkpoints/i3d_r50_256p_32x2x1_100e_kinetics400_rgb_20200801-7d9f44de.pth PATH_TO_LONG_VIDEO tools/data/kinetics/label_map_k400.txt PATH_TO_SAVED_VIDEO \
+     --label-color 255 255 0
+   ```
+
+5. Predict different labels in a long video by using a I3D model on gpu and save the results as a `json` file
+
+   ```shell
+   python demo/long_video_demo.py configs/recognition/i3d/i3d_r50_video_inference_32x2x1_100e_kinetics400_rgb.py \
+     checkpoints/i3d_r50_256p_32x2x1_100e_kinetics400_rgb_20200801-7d9f44de.pth PATH_TO_LONG_VIDEO tools/data/kinetics/label_map_k400.txt ./results.json
+   ```
+
+## SpatioTemporal Action Detection Webcam Demo
+
+We provide a demo script to implement real-time spatio-temporal action detection from a web camera.
+
+```shell
+python demo/webcam_demo_spatiotemporal_det.py \
+    [--config ${SPATIOTEMPORAL_ACTION_DETECTION_CONFIG_FILE}] \
+    [--checkpoint ${SPATIOTEMPORAL_ACTION_DETECTION_CHECKPOINT}] \
+    [--action-score-thr ${ACTION_DETECTION_SCORE_THRESHOLD}] \
+    [--det-config ${HUMAN_DETECTION_CONFIG_FILE}] \
+    [--det-checkpoint ${HUMAN_DETECTION_CHECKPOINT}] \
+    [--det-score-thr ${HUMAN_DETECTION_SCORE_THRESHOLD}] \
+    [--input-video] ${INPUT_VIDEO} \
+    [--label-map ${LABEL_MAP}] \
+    [--device ${DEVICE}] \
+    [--output-fps ${OUTPUT_FPS}] \
+    [--out-filename ${OUTPUT_FILENAME}] \
+    [--show] \
+    [--display-height] ${DISPLAY_HEIGHT} \
+    [--display-width] ${DISPLAY_WIDTH} \
+    [--predict-stepsize ${PREDICT_STEPSIZE}] \
+    [--clip-vis-length] ${CLIP_VIS_LENGTH}
+```
+
+Optional arguments:
+
+- `SPATIOTEMPORAL_ACTION_DETECTION_CONFIG_FILE`: The spatiotemporal action detection config file path.
+- `SPATIOTEMPORAL_ACTION_DETECTION_CHECKPOINT`: The spatiotemporal action detection checkpoint path or URL.
+- `ACTION_DETECTION_SCORE_THRESHOLD`: The score threshold for action detection. Default: 0.4.
+- `HUMAN_DETECTION_CONFIG_FILE`: The human detection config file path.
+- `HUMAN_DETECTION_CHECKPOINT`: The human detection checkpoint URL.
+- `HUMAN_DETECTION_SCORE_THRE`: The score threshold for human detection. Default: 0.9.
+- `INPUT_VIDEO`: The webcam id or video path of the source. Default: `0`.
+- `LABEL_MAP`: The label map used. Default: `tools/data/ava/label_map.txt`.
+- `DEVICE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`.  Default: `cuda:0`.
+- `OUTPUT_FPS`: The FPS of demo video output. Default: 15.
+- `OUTPUT_FILENAME`: Path to the output file which is a video format. Default: None.
+- `--show`: Whether to show predictions with `cv2.imshow`.
+- `DISPLAY_HEIGHT`: The height of the display frame. Default: 0.
+- `DISPLAY_WIDTH`: The width of the display frame. Default: 0. If `DISPLAY_HEIGHT <= 0 and DISPLAY_WIDTH <= 0`, the display frame and input video share the same shape.
+- `PREDICT_STEPSIZE`: Make a prediction per N frames. Default: 8.
+- `CLIP_VIS_LENGTH`: The number of the draw frames for each clip. In other words, for each clip, there are at most `CLIP_VIS_LENGTH` frames to be draw around the keyframe. DEFAULT: 8.
+
+Tips to get a better experience for webcam demo:
+
+- How to choose `--output-fps`?
+
+  - `--output-fps` should be almost equal to read thread fps.
+  - Read thread fps is printed by logger in format `DEBUG:__main__:Read Thread: {duration} ms, {fps} fps`
+
+- How to choose `--predict-stepsize`?
+
+  - It's related to how to choose human detector and spatio-temporval model.
+  - Overall, the duration of read thread for each task should be greater equal to that of model inference.
+  - The durations for read/inference are both printed by logger.
+  - Larger `--predict-stepsize` leads to larger duration for read thread.
+  - In order to fully take the advantage of computation resources, decrease the value of `--predict-stepsize`.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` .
+
+1. Use the Faster RCNN as the human detector, SlowOnly-8x8-R101 as the action detector. Making predictions per 40 frames, and FPS of the output is 20. Show predictions with `cv2.imshow`.
+
+```shell
+python demo/webcam_demo_spatiotemporal_det.py \
+    --input-video 0 \
+    --config configs/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb.py \
+    --checkpoint https://download.openmmlab.com/mmaction/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb_20201217-16378594.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --det-score-thr 0.9 \
+    --action-score-thr 0.5 \
+    --label-map tools/data/ava/label_map.txt \
+    --predict-stepsize 40 \
+    --output-fps 20 \
+    --show
+```
+
+## Skeleton-based Action Recognition Demo
+
+We provide a demo script to predict the skeleton-based action recognition result using a single video.
+
+```shell
+python demo/demo_skeleton.py ${VIDEO_FILE} ${OUT_FILENAME} \
+    [--config ${SKELETON_BASED_ACTION_RECOGNITION_CONFIG_FILE}] \
+    [--checkpoint ${SKELETON_BASED_ACTION_RECOGNITION_CHECKPOINT}] \
+    [--det-config ${HUMAN_DETECTION_CONFIG_FILE}] \
+    [--det-checkpoint ${HUMAN_DETECTION_CHECKPOINT}] \
+    [--det-score-thr ${HUMAN_DETECTION_SCORE_THRESHOLD}] \
+    [--pose-config ${HUMAN_POSE_ESTIMATION_CONFIG_FILE}] \
+    [--pose-checkpoint ${HUMAN_POSE_ESTIMATION_CHECKPOINT}] \
+    [--label-map ${LABEL_MAP}] \
+    [--device ${DEVICE}] \
+    [--short-side] ${SHORT_SIDE}
+```
+
+Optional arguments:
+
+- `SKELETON_BASED_ACTION_RECOGNITION_CONFIG_FILE`: The skeleton-based action recognition config file path.
+- `SKELETON_BASED_ACTION_RECOGNITION_CHECKPOINT`: The skeleton-based action recognition checkpoint path or URL.
+- `HUMAN_DETECTION_CONFIG_FILE`: The human detection config file path.
+- `HUMAN_DETECTION_CHECKPOINT`: The human detection checkpoint URL.
+- `HUMAN_DETECTION_SCORE_THRE`: The score threshold for human detection. Default: 0.9.
+- `HUMAN_POSE_ESTIMATION_CONFIG_FILE`: The human pose estimation config file path (trained on COCO-Keypoint).
+- `HUMAN_POSE_ESTIMATION_CHECKPOINT`: The human pose estimation checkpoint URL (trained on COCO-Keypoint).
+- `LABEL_MAP`: The label map used. Default: `tools/data/ava/label_map.txt`.
+- `DEVICE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`.  Default: `cuda:0`.
+- `SHORT_SIDE`: The short side used for frame extraction. Default: 480.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` .
+
+1. Use the Faster RCNN as the human detector, HRNetw32 as the pose estimator, PoseC3D-NTURGB+D-120-Xsub-keypoint as the skeleton-based action recognizer.
+
+```shell
+python demo/demo_skeleton.py demo/ntu_sample.avi demo/skeleton_demo.mp4 \
+    --config configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py \
+    --checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint/slowonly_r50_u48_240e_ntu120_xsub_keypoint-6736b03f.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --det-score-thr 0.9 \
+    --pose-config demo/hrnet_w32_coco_256x192.py \
+    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
+    --label-map tools/data/skeleton/label_map_ntu120.txt
+```
+
+2. Use the Faster RCNN as the human detector, HRNetw32 as the pose estimator, STGCN-NTURGB+D-60-Xsub-keypoint as the skeleton-based action recognizer.
+
+```shell
+python demo/demo_skeleton.py demo/ntu_sample.avi demo/skeleton_demo.mp4 \
+    --config configs/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint.py \
+    --checkpoint https://download.openmmlab.com/mmaction/skeleton/stgcn/stgcn_80e_ntu60_xsub_keypoint/stgcn_80e_ntu60_xsub_keypoint-e7bb9653.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --det-score-thr 0.9 \
+    --pose-config demo/hrnet_w32_coco_256x192.py \
+    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
+    --label-map tools/data/skeleton/label_map_ntu120.txt
+```
+
+## Video Structuralize Demo
+
+We provide a demo script to to predict the skeleton-based and rgb-based action recognition and spatio-temporal action detection result using a single video.
+
+```shell
+python demo/demo_video_structuralize.py
+    [--rgb-stdet-config ${RGB_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CONFIG_FILE}] \
+    [--rgb-stdet-checkpoint ${RGB_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CHECKPOINT}] \
+    [--skeleton-stdet-checkpoint ${SKELETON_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CHECKPOINT}] \
+    [--det-config ${HUMAN_DETECTION_CONFIG_FILE}] \
+    [--det-checkpoint ${HUMAN_DETECTION_CHECKPOINT}] \
+    [--pose-config ${HUMAN_POSE_ESTIMATION_CONFIG_FILE}] \
+    [--pose-checkpoint ${HUMAN_POSE_ESTIMATION_CHECKPOINT}] \
+    [--skeleton-config ${SKELETON_BASED_ACTION_RECOGNITION_CONFIG_FILE}] \
+    [--skeleton-checkpoint ${SKELETON_BASED_ACTION_RECOGNITION_CHECKPOINT}] \
+    [--rgb-config ${RGB_BASED_ACTION_RECOGNITION_CONFIG_FILE}] \
+    [--rgb-checkpoint ${RGB_BASED_ACTION_RECOGNITION_CHECKPOINT}] \
+    [--use-skeleton-stdet ${USE_SKELETON_BASED_SPATIO_TEMPORAL_DETECTION_METHOD}] \
+    [--use-skeleton-recog ${USE_SKELETON_BASED_ACTION_RECOGNITION_METHOD}] \
+    [--det-score-thr ${HUMAN_DETECTION_SCORE_THRE}] \
+    [--action-score-thr ${ACTION_DETECTION_SCORE_THRE}] \
+    [--video ${VIDEO_FILE}] \
+    [--label-map-stdet ${LABEL_MAP_FOR_SPATIO_TEMPORAL_ACTION_DETECTION}] \
+    [--device ${DEVICE}] \
+    [--out-filename ${OUTPUT_FILENAME}] \
+    [--predict-stepsize ${PREDICT_STEPSIZE}] \
+    [--output-stepsize ${OUTPU_STEPSIZE}] \
+    [--output-fps ${OUTPUT_FPS}] \
+    [--cfg-options]
+```
+
+Optional arguments:
+
+- `RGB_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CONFIG_FILE`: The rgb-based spatio temoral action detection config file path.
+- `RGB_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CHECKPOINT`: The rgb-based spatio temoral action detection checkpoint path or URL.
+- `SKELETON_BASED_SPATIO_TEMPORAL_ACTION_DETECTION_CHECKPOINT`: The skeleton-based spatio temoral action detection checkpoint path or URL.
+- `HUMAN_DETECTION_CONFIG_FILE`: The human detection config file path.
+- `HUMAN_DETECTION_CHECKPOINT`: The human detection checkpoint URL.
+- `HUMAN_POSE_ESTIMATION_CONFIG_FILE`: The human pose estimation config file path (trained on COCO-Keypoint).
+- `HUMAN_POSE_ESTIMATION_CHECKPOINT`: The human pose estimation checkpoint URL (trained on COCO-Keypoint).
+- `SKELETON_BASED_ACTION_RECOGNITION_CONFIG_FILE`: The skeleton-based action recognition config file path.
+- `SKELETON_BASED_ACTION_RECOGNITION_CHECKPOINT`: The skeleton-based action recognition checkpoint path or URL.
+- `RGB_BASED_ACTION_RECOGNITION_CONFIG_FILE`: The rgb-based action recognition config file path.
+- `RGB_BASED_ACTION_RECOGNITION_CHECKPOINT`: The rgb-based action recognition checkpoint path or URL.
+- `USE_SKELETON_BASED_SPATIO_TEMPORAL_DETECTION_METHOD`: Use skeleton-based spatio temporal action detection method.
+- `USE_SKELETON_BASED_ACTION_RECOGNITION_METHOD`: Use skeleton-based action recognition method.
+- `HUMAN_DETECTION_SCORE_THRE`: The score threshold for human detection. Default: 0.9.
+- `ACTION_DETECTION_SCORE_THRE`: The score threshold for action detection. Default: 0.4.
+- `LABEL_MAP_FOR_SPATIO_TEMPORAL_ACTION_DETECTION`: The label map for spatio temporal action detection used. Default: `tools/data/ava/label_map.txt`.
+- `LABEL_MAP`: The label map for action recognition. Default: `tools/data/kinetics/label_map_k400.txt`.
+- `DEVICE`: Type of device to run the demo. Allowed values are cuda device like `cuda:0` or `cpu`.  Default: `cuda:0`.
+- `OUTPUT_FILENAME`: Path to the output file which is a video format. Default: `demo/test_stdet_recognition_output.mp4`.
+- `PREDICT_STEPSIZE`: Make a prediction per N frames.  Default: 8.
+- `OUTPUT_STEPSIZE`: Output 1 frame per N frames in the input video. Note that `PREDICT_STEPSIZE % OUTPUT_STEPSIZE == 0`. Default: 1.
+- `OUTPUT_FPS`: The FPS of demo video output. Default: 24.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` .
+
+1. Use the Faster RCNN as the human detector, HRNetw32 as the pose estimator, PoseC3D as the skeleton-based action recognizer and the skeleton-based spatio temporal action detector. Making action detection predictions per 8 frames, and output 1 frame per 1 frame to the output video. The FPS of the output video is 24.
+
+```shell
+python demo/demo_video_structuralize.py
+    --skeleton-stdet-checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/posec3d_ava.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --pose-config demo/hrnet_w32_coco_256x192.py
+    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/
+    hrnet_w32_coco_256x192-c78dce93_20200708.pth \
+    --skeleton-config configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py \
+    --skeleton-checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/
+    posec3d_k400.pth \
+    --use-skeleton-stdet \
+    --use-skeleton-recog \
+    --label-map-stdet tools/data/ava/label_map.txt \
+    --label-map tools/data/kinetics/label_map_k400.txt
+```
+
+2. Use the Faster RCNN as the human detector, TSN-R50-1x1x3 as the rgb-based action recognizer, SlowOnly-8x8-R101 as the rgb-based spatio temporal action detector. Making action detection predictions per 8 frames, and output 1 frame per 1 frame to the output video. The FPS of the output video is 24.
+
+```shell
+python demo/demo_video_structuralize.py
+    --rgb-stdet-config configs/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb.py \
+    --rgb-stdet-checkpoint  https://download.openmmlab.com/mmaction/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb_20201217-16378594.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --rgb-config configs/recognition/tsn/
+    tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+    --rgb-checkpoint https://download.openmmlab.com/mmaction/recognition/
+    tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/
+    tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+    --label-map-stdet tools/data/ava/label_map.txt \
+    --label-map tools/data/kinetics/label_map_k400.txt
+```
+
+3. Use the Faster RCNN as the human detector, HRNetw32 as the pose estimator, PoseC3D as the skeleton-based action recognizer, SlowOnly-8x8-R101 as the rgb-based spatio temporal action detector. Making action detection predictions per 8 frames, and output 1 frame per 1 frame to the output video. The FPS of the output video is 24.
+
+```shell
+python demo/demo_video_structuralize.py
+    --rgb-stdet-config configs/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb.py \
+    --rgb-stdet-checkpoint  https://download.openmmlab.com/mmaction/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb_20201217-16378594.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --pose-config demo/hrnet_w32_coco_256x192.py
+    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/
+    hrnet_w32_coco_256x192-c78dce93_20200708.pth \
+    --skeleton-config configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py \
+    --skeleton-checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/
+    posec3d_k400.pth \
+    --use-skeleton-recog \
+    --label-map-stdet tools/data/ava/label_map.txt \
+    --label-map tools/data/kinetics/label_map_k400.txt
+```
+
+4. Use the Faster RCNN as the human detector, HRNetw32 as the pose estimator, TSN-R50-1x1x3 as the rgb-based action recognizer, PoseC3D as the skeleton-based spatio temporal action detector. Making action detection predictions per 8 frames, and output 1 frame per 1 frame to the output video. The FPS of the output video is 24.
+
+```shell
+python demo/demo_video_structuralize.py
+    --skeleton-stdet-checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/posec3d_ava.pth \
+    --det-config demo/faster_rcnn_r50_fpn_2x_coco.py \
+    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
+    --pose-config demo/hrnet_w32_coco_256x192.py
+    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/
+    hrnet_w32_coco_256x192-c78dce93_20200708.pth \
+    --skeleton-config configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py \
+    --rgb-config configs/recognition/tsn/
+    tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py \
+    --rgb-checkpoint https://download.openmmlab.com/mmaction/recognition/
+    tsn/tsn_r50_1x1x3_100e_kinetics400_rgb/
+    tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth \
+    --use-skeleton-stdet \
+    --label-map-stdet tools/data/ava/label_map.txt \
+    --label-map tools/data/kinetics/label_map_k400.txt
+```
+
+## Audio Demo
+
+Demo script to predict the audio-based action recognition using a single audio feature.
+
+The script `extract_audio.py` can be used to extract audios from videos and the script `build_audio_features.py` can be used to extract the audio features.
+
+```shell
+python demo/demo_audio.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${AUDIO_FILE} {LABEL_FILE} [--device ${DEVICE}]
+```
+
+Optional arguments:
+
+- `DEVICE`: Type of device to run the demo. Allowed values are cuda devices like `cuda:0` or `cpu`. If not specified, it will be set to `cuda:0`.
+
+Examples:
+
+Assume that you are located at `$MMACTION2` and have already downloaded the checkpoints to the directory `checkpoints/`,
+or use checkpoint url from `configs/` to directly load the corresponding checkpoint, which will be automatically saved in `$HOME/.cache/torch/checkpoints`.
+
+1. Recognize an audio file as input by using a tsn model on cuda by default.
+
+   ```shell
+   python demo/demo_audio.py \
+       configs/recognition_audio/resnet/tsn_r18_64x1x1_100e_kinetics400_audio_feature.py \
+       https://download.openmmlab.com/mmaction/recognition/audio_recognition/tsn_r18_64x1x1_100e_kinetics400_audio_feature/tsn_r18_64x1x1_100e_kinetics400_audio_feature_20201012-bf34df6c.pth \
+       audio_feature.npy label_map_k400.txt
+   ```

openmmlab_test/mmaction2-0.24.1/demo/demo.ipynb

+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {
+    "pycharm": {
+     "is_executing": false
+    }
+   },
+   "outputs": [],
+   "source": [
+    "from mmaction.apis import init_recognizer, inference_recognizer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "pycharm": {
+     "is_executing": false
+    }
+   },
+   "outputs": [],
+   "source": [
+    "config_file = '../configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py'\n",
+    "# download the checkpoint from model zoo and put it in `checkpoints/`\n",
+    "checkpoint_file = '../checkpoints/tsn_r50_1x1x3_100e_kinetics400_rgb_20200614-e508be42.pth'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {
+    "pycharm": {
+     "is_executing": false
+    }
+   },
+   "outputs": [],
+   "source": [
+    "# build the model from a config file and a checkpoint file\n",
+    "model = init_recognizer(config_file, checkpoint_file, device='cpu')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {
+    "pycharm": {
+     "is_executing": false
+    }
+   },
+   "outputs": [],
+   "source": [
+    "# test a single video and show the result:\n",
+    "video = 'demo.mp4'\n",
+    "label = '../tools/data/kinetics/label_map_k400.txt'\n",
+    "results = inference_recognizer(model, video)\n",
+    "\n",
+    "labels = open(label).readlines()\n",
+    "labels = [x.strip() for x in labels]\n",
+    "results = [(labels[k[0]], k[1]) for k in results]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {
+    "collapsed": false,
+    "jupyter": {
+     "outputs_hidden": false
+    },
+    "pycharm": {
+     "is_executing": false,
+     "name": "#%%\n"
+    }
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "arm wrestling:  29.61644\n",
+      "rock scissors paper:  10.754839\n",
+      "shaking hands:  9.9084\n",
+      "clapping:  9.189912\n",
+      "massaging feet:  8.305307\n"
+     ]
+    }
+   ],
+   "source": [
+    "# show the results\n",
+    "for result in results:\n",
+    "    print(f'{result[0]}: ', result[1])"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.7.4"
+  },
+  "pycharm": {
+   "stem_cell": {
+    "cell_type": "raw",
+    "metadata": {
+     "collapsed": false
+    },
+    "source": []
+   }
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}