Skip to content

Latest commit

 

History

History
71 lines (56 loc) · 2.82 KB

reproduce.md

File metadata and controls

71 lines (56 loc) · 2.82 KB

Instructions to reproduce our results using pretrained models

Here we provide a step-step instruction to reproduce our results using our docker image.

I: The zip files of pretrained models should be decompressed and move to their folders:

  • Conv3D/pretrained.zip needs to be decompressed and put in Conv3D/ folder in our code repo. (contains pretrained RGB model using Chinese Sign Language dataset)
  • Conv3D/final_models_finetuned.zip needs to be decompressed and put in Conv3D/ folder in our code repo. (contains pretrained our final models of all modalities)
  • SL-GCN/27_2.zip contains models trained using training set only and needs to be decompressed and put under SL-GCN/final_models/ in our code repo.
  • SL-GCN/27_2_finetuned.zip contains models trained using training set and validation set which needs to be decompressed and put under SL-GCN/final_models/ in our code repo.
  • SSTCN/T_Pose_model_final.pth do not need decompressing and can be put under SSTCN/ folder directly.

II: Nvidia docker image:

Our docker image can be load using the following command:

sudo docker run -it --gpus all -v path_to_your_data:/home/smilelab_slr/cvpr2021_allcode/shared_data cvpr2021cha_code /bin/bash

III: Procedues to reproduce our results

Here I provide a brief pipeline to help reproduce our results, detailed instruction can be found in our code repo:

A. For RGB track:

  1. Please test SL-GCN using the following script. All the config files in SL-GCN/config/test and SL-GCN/config/test_finetuned need to be tested.
python main.py --config /path/to/config/file
  1. Please test the other modalities using:
cd Conv3D
python Sign_Isolated_Conv3D_clip_test.py
python Sign_Isolated_Conv3D_flow_clip_test.py
cd SSTCN
python test.py

The test results are saved as .pkl files.

  1. Ensemble the final results following the instructions in ensemble/ folder.
  • ensemble gcn multi-stream results first.
  • ensemble rgb results using ensemble_multimodal_rgb.py

For RGB-D track:

  1. Please test SL-GCN using the following script. All the config files in SL-GCN/config/test and SL-GCN/config/test_finetuned need to be tested.
python main.py --config /path/to/config/file
  1. Please 'cd' to each folder and test the pretrained models using the following scripts:
cd Conv3D
python Sign_Isolated_Conv3D_clip_test.py
python Sign_Isolated_Conv3D_flow_clip_test.py
python Sign_Isolated_Conv3D_hha_clip_mask_test.py
python /Sign_Isolated_Conv3D_depth_flow_clip_test.py
cd SSTCN
python test.py

The test results are saved as .pkl files.

  1. Ensemble the final results following the instructions in ensemble/ folder.
  • ensemble gcn multi-stream results first.
  • ensemble rgbd results using ensemble_multimodal_rgbd.py

Please note that you need to rename and copy all the generated .pkl files to the ensemble folder.