Introduced 3D CBAM Attention to the PoseC3D in MMAction2
本Git只提供了code,实际要到https://github.com/open-mmlab/mmaction2 这个git页去clone跑,新py文件替换原本文件
配置教程:https://blog.csdn.net/WhiffeYF/article/details/120556253
针对于一个数据集
先在data下 wget pkl file (annotation file)
再train: 在mmaction2下跑./tools/dist_train.sh ./configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_ucf101_split1_keypoint.py 2 --validate
再test :在mmaction2下跑 ./tools/dist_test.sh ./configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_ucf101_split1_keypoint.py ./work_dirs/posec3d_iclr/slowonly_kinetics400_pretrained_r50_u48_120e_ucf101_split1_keypoint/best_top1_acc_epoch_11.pth 2 --eval top_k_accuracy
路径都是一样的
ntu60的keypoint和limb各自分开都可以用
topkaccuracy全部公用
benchmark是榜单的意思 (paper with code上找)
heatmap visualization直接在jupyter里跑,下载好pickle就行,放到相应路径
ntu60subtrain是train,val是测试
NTU60: ./tools/dist_train.sh ./configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint.py 4 --validate
./tools/dist_test.sh ./configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint.py ./work_dirs/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint/best_top1_acc_epoch_11.pth 2 --eval top_k_accuracy
HMDB: ./tools/dist_train.sh ./configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_hmdb51_split1_keypoint.py 2 --validate
./tools/dist_test.sh ./configs/skeleton/posec3d/slowonly_kinetics400_pretrained_r50_u48_120e_hmdb51_split1_keypoint.py ./work_dirs/posec3d_iclr/slowonly_kinetics400_pretrained_r50_u48_120e_hmdb51_split1_keypoint/best_top1_acc_epoch_12.pth 2 --eval top_k_accuracy
为了保持结构的一致性,我把configs里的ucf和hmdb里面stage_blocks改成了(4,6,3). 总共只有三个stage