Official Implementation for CVAR

Environment

• python=3.7
• PyTorch=1.6.0

Original Github Repo from Yuexi Zhang

Dataset:

Northwestern-UCLA Multiview Action 3D Dataset(NUCLA)
List for Cross View setup.

Training:

Almost all hyperparameters included as default values(Now only support NUCLA/Cross-View/Single/Gumbel/Re-Weighted DYAN(RHdyan)). \

1. DIR-Dictionary(1145MiB GRAM)

       python train_DIR_D_NUCLA.py --path_list /dir/to/data  --modelRoot /dir/to/save --bs 64 --lam_f 0.01 --gumbel_thresh 0.5 --gpu_id 0  > /dir/to/save/DIR_D_NUCLA.log

   Or different loss()

      python train_DIR_D_NUCLA_lossR_B.py --path_list /dir/to/data  --modelRoot /dir/to/save --bs 64 --lam_f 0.01 --gumbel_thresh 0.5 --gpu_id 0  > /dir/to/save/DIR_D_NUCLA.log

2. DIR-Classification(1107MiB GRAM),Pretrained Dictionary from step 1

       python train_DIR_cls_noCL_NUCLA.py --path_list /dir/to/data  --modelRoot /dir/to/save --pretrain /path/to/model --bs 64 --lam_f 0.01 --gumbel_thresh 0.5 --gpu_id 0  > /dir/to/save/DIR_cls_noCL_NUCLA.log

3. DIR-Contrastive Learning(Not Supported Yet)

       python train_DIR_cls_wiCL_NUCLA.py --path_list /dir/to/data --modelRoot /dir/to/save --pretrain /path/to/model

4. DIR-Finetune(1023MiB GRAM)(Not Supported Yet)

       python train_DIR_cls_ft_NUCLA.py --modelRoot /dir/to/save --path_list /dir/to/data --pretrain /path/to/model

Human Interaction Recognition experiment can be found:

https://github.com/DanLuoNEU/Skeleton-based_Interactive_Action_Recognition_via_Dynamics