search

fmthoker / skeleton-contrast

41 stars 5 forks source link

Has anyone reproduced the reported performance in the paper? #7

Closed dingli-dean closed 2 years ago

dingli-dean commented 2 years ago

I have tried to reproduce the performance, but limited success was achieved. Only got 67.20% acc with intra-skeleton contrast of graph-based network, while the corresponding performance is 72.5% in table 6 of the paper.

fmthoker commented 2 years ago

can you upload the training config and logs?

dingli-dean commented 2 years ago

Self-supervised training config: {'graph_args': {'labeling_mode': 'spatial'}, 'num_class': 128, 'num_person': 2, 'num_point': 25}

{'data_path': '/media/liding/dataset/NTU/data_Skeleton_Contrast/NTU-RGB-D-60-AGCN/xview/train_data_joint.npy', 'input_representation': 'graph-based', 'input_size': 64, 'l_ratio': [0.1, 1], 'num_frame_path': '/media/liding/dataset/NTU/data_Skeleton_Contrast/NTU-RGB-D-60-AGCN/xview/train_num_frame.npy'}{'batch_size': 256, 'checkpoint_path': '/media/liding/experiments/sk/exp_g-2/ntu_60_cross_view/intraskeleton_graph_based', 'cos': False, 'dev': 33, 'epochs': 451, 'gpu': 0, 'lr': 0.01, 'mlp': True, 'moco_dim': 128, 'moco_k': 16384, 'moco_m': 0.999, 'moco_t': 0.07, 'momentum': 0.9, 'pre_dataset': 'ntu60', 'print_freq': 10, 'protocol': 'cross_view', 'resume': '', 'schedule': [351], 'seed': None, 'skeleton_representation': 'graph-based', 'start_epoch': 0, 'weight_decay': 0.0001, 'workers': 32}

Self-supervised training tensorboard log is: g_1_ssl_traning

Linear evaluation config is: -------agcn_model_args------- {'graph_args': {'labeling_mode': 'spatial'}, 'num_class': 60, 'num_person': 2, 'num_point': 25} -------train_feeder_args------- {'data_path': '/home/liding/dataset/NTU/data_Skeleton_Contrast/NTU-RGB-D-60-AGCN/xview/train_data_joint.npy', 'input_representation': 'graph-based', 'input_size': 64, 'l_ratio': [0.5, 1.0], 'label_path': '/home/liding/dataset/NTU/data_Skeleton_Contrast/NTU-RGB-D-60-AGCN/xview/train_label.pkl', 'num_frame_path': '/home/liding/dataset/NTU/data_Skeleton_Contrast/NTU-RGB-D-60-AGCN/xview/train_num_frame.npy'} -------exp_args------- {'batch_size': 1024, 'checkpoint_path': '/media/liding/experiments/sk/exp_g-2/ntu_60_cross_view/intraskeleton_graph_based/action_cls_joint', 'dev': 31, 'epochs': 80, 'finetune_dataset': 'ntu60', 'finetune_skeleton_representation': 'graph-based', 'gpu': 0, 'lr': 0.1, 'momentum': 0.9, 'pretrain_skeleton_representation': 'graph-based', 'pretrained': '/media/liding/experiments/sk/exp_g-2/ntu_60_cross_view/intraskeleton_graph_based/checkpoint_0450.pth.tar', 'print_freq': 10, 'protocol': 'cross_view', 'resume': '', 'schedule': [50, 70], 'seed': None, 'start_epoch': 0, 'weight_decay': 0.0, 'workers': 32}

dingli-dean commented 2 years ago

It is the batch_size in linear classifier training that hinders the reproduction of experiment. Do not modify defaut number in .sh file. Set the batch_size == 64.