How can I train a gesture recognition model with a dynamic dataset?

[MJCC-wht]

I want to use stgcn++, so I mean which config file I can use? Or I need to create my own config file?

[karryor]

Hello, have you trained gesture recognition with stgcn++? Can you share it?

[MJCC-wht]

Hello, have you trained gesture recognition with stgcn++? Can you share it?

I just train simply, the config file content is below:

graph = 'handmp'
modality = 'j'

model = dict(
type='RecognizerGCN',
backbone=dict(
type='STGCN',
in_channels=2,
gcn_adaptive='init',
gcn_with_res=True,
tcn_type='mstcn',
num_stages=6,
down_stages=[6],
inflate_stages=[6],
graph_cfg=dict(layout=graph, mode='spatial')),
cls_head=dict(type='GCNHead', num_classes=2, in_channels=128))

dataset_type = 'GestureDataset'
ann_file = 'data/labdata/smoke_output_final.pkl'
load_from = 'demo/hagrid.pth'
train_pipeline = [
dict(type='PreNormalize2D', threshold=0, mode='auto'),
dict(type='GenSkeFeat', dataset=graph, feats=[modality]),
dict(type='UniformSample', clip_len=10, num_clips=1),
dict(type='PoseDecode'),
dict(type='FormatGCNInput', num_person=1),
dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
dict(type='ToTensor', keys=['keypoint'])
]
val_pipeline = [
dict(type='PreNormalize2D', threshold=0, mode='auto'),
dict(type='GenSkeFeat', dataset=graph, feats=[modality]),
dict(type='UniformSample', clip_len=10, num_clips=1),
dict(type='PoseDecode'),
dict(type='FormatGCNInput', num_person=1),
dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]),
dict(type='ToTensor', keys=['keypoint'])
]
test_pipeline = [
dict(type='PreNormalize2D', threshold=0, mode='auto'),
dict(type='GenSkeFeat', dataset=graph, feats=[modality]),
dict(type='UniformSample', clip_len=10, num_clips=1),
dict(type='PoseDecode'),
dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')),
val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'),
test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val'))

optimizer

optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True)
optimizer_config = dict(grad_clip=None)

learning policy

lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False)
total_epochs = 25
checkpoint_config = dict(interval=1)
evaluation = dict(interval=1, metrics=['top_k_accuracy'])
log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')])

runtime settings

log_level = 'INFO'
work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j'

then just use the script bash file to train

[karryor]

Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file.

…

------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" ***@***.***>; 发送时间: 2024年1月16日(星期二) 下午2:55 ***@***.***>; ***@***.******@***.***>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[MJCC-wht]

Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file.
…
------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月16日(星期二) 下午2:55 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

Sorry, I didn't. Maybe you need to try it on your own.

[karryor]

OK，thank you!

…

------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" ***@***.***>; 发送时间: 2024年1月17日(星期三) 下午4:58 ***@***.***>; ***@***.******@***.***>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file. … ------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月16日(星期二) 下午2:55 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***> Sorry, I didn't. Maybe you need to try it on your own. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[karryor]

Hello, can you share the 40 categories of gesture recognition and corresponding gesture images， thank you？

…

------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" ***@***.***>; 发送时间: 2024年1月17日(星期三) 下午4:58 ***@***.***>; ***@***.******@***.***>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file. … ------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月16日(星期二) 下午2:55 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***> Sorry, I didn't. Maybe you need to try it on your own. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[MJCC-wht]

Hello, can you share the 40 categories of gesture recognition and corresponding gesture images， thank you？
…
------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月17日(星期三) 下午4:58 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file. … ------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月16日(星期二) 下午2:55 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> Sorry, I didn't. Maybe you need to try it on your own. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

Sorry, I use my own dataset which has only 2 categories, so I can't provide the images

[karryor]

OK, thank you.

…

------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" ***@***.***>; 发送时间: 2024年1月19日(星期五) 中午11:12 ***@***.***>; ***@***.******@***.***>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, can you share the 40 categories of gesture recognition and corresponding gesture images， thank you？ … ------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月17日(星期三) 下午4:58 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, thank you for your reply. There is one more question, I have learned that STGCN network classifs behaviors for consecutive frames. Did you add tracking network? I didn't see it in the configuration file. … ------------------ 原始邮件 ------------------ 发件人: "kennymckormick/pyskl" @.>; 发送时间: 2024年1月16日(星期二) 下午2:55 @.>; @.@.>; 主题: Re: [kennymckormick/pyskl] How can I train a gesture recognition model with a dynamic dataset? (Issue #213) Hello, have you trained gesture recognition with stgcn++? Can you share it? I just train simply, the config file content is below: graph = 'handmp' modality = 'j' model = dict( type='RecognizerGCN', backbone=dict( type='STGCN', in_channels=2, gcn_adaptive='init', gcn_with_res=True, tcn_type='mstcn', num_stages=6, down_stages=[6], inflate_stages=[6], graph_cfg=dict(layout=graph, mode='spatial')), cls_head=dict(type='GCNHead', num_classes=2, in_channels=128)) dataset_type = 'GestureDataset' ann_file = 'data/labdata/smoke_output_final.pkl' load_from = 'demo/hagrid.pth' train_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] val_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=1), dict(type='Collect', keys=['keypoint', 'label'], meta_keys=[]), dict(type='ToTensor', keys=['keypoint']) ] test_pipeline = [ dict(type='PreNormalize2D', threshold=0, mode='auto'), dict(type='GenSkeFeat', dataset=graph, feats=[modality]), dict(type='UniformSample', clip_len=10, num_clips=1), dict(type='PoseDecode'), dict(type='FormatGCNInput', 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, pipeline=train_pipeline, split='labdata_train')), val=dict(type=dataset_type, ann_file=ann_file, pipeline=val_pipeline, split='labdata_val'), test=dict(type=dataset_type, ann_file=ann_file, pipeline=test_pipeline, split='labdata_val')) optimizer optimizer = dict(type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0005, nesterov=True) optimizer_config = dict(grad_clip=None) learning policy lr_config = dict(policy='CosineAnnealing', min_lr=0, by_epoch=False) total_epochs = 25 checkpoint_config = dict(interval=1) evaluation = dict(interval=1, metrics=['top_k_accuracy']) log_config = dict(interval=1, hooks=[dict(type='TextLoggerHook')]) runtime settings log_level = 'INFO' work_dir = './work_dirs/stgcn++/stgcn++_labdata_hrnet/hand_j' then just use the script bash file to train — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> Sorry, I didn't. Maybe you need to try it on your own. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> Sorry, I use my own dataset which has only 2 categories, so I can't provide the images — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>