MS-G3D

PyTorch implementation of "Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition", CVPR 2020 Oral.

[PDF][Demo][Abstract/Supp]

Dependencies

• Python >= 3.6
• PyTorch >= 1.2.0
• NVIDIA Apex (auto mixed precision training)
• PyYAML, tqdm, tensorboardX

Data Preparation

Disk usage warning: after preprocessing, the total sizes of datasets are around 38GB, 77GB, 63GB for NTU RGB+D 60, NTU RGB+D 120, and Kinetics 400, respectively. The raw/intermediate sizes may be larger.

Download Datasets

There are 3 datasets to download:

• NTU RGB+D 60 Skeleton
• NTU RGB+D 120 Skeleton
• Kinetics 400 Skeleton

NTU RGB+D 60 and 120

1. Request dataset here: http://rose1.ntu.edu.sg/Datasets/actionRecognition.asp

2. Download the skeleton-only datasets:

   • nturgbd_skeletons_s001_to_s017.zip (NTU RGB+D 60)
   • nturgbd_skeletons_s018_to_s032.zip (NTU RGB+D 120, on top of NTU RGB+D 60)
   • Total size should be 5.8GB + 4.5GB.

3. Download missing skeletons lookup files from the authors' GitHub repo:

   • NTU RGB+D 60 Missing Skeletons: wget https://raw.githubusercontent.com/shahroudy/NTURGB-D/master/Matlab/NTU_RGBD_samples_with_missing_skeletons.txt

   • NTU RGB+D 120 Missing Skeletons: wget https://raw.githubusercontent.com/shahroudy/NTURGB-D/master/Matlab/NTU_RGBD120_samples_with_missing_skeletons.txt

   • Remember to remove the first few lines of text in these files!

Kinetics Skeleton 400

1. Download dataset from ST-GCN repo: https://github.com/yysijie/st-gcn/blob/master/OLD_README.md#kinetics-skeleton
2. This might be useful if you want to wget the dataset from Google Drive

Data Preprocessing

Directory Structure

Put downloaded data into the following directory structure:

- data/
  - kinetics_raw/
    - kinetics_train/
      ...
    - kinetics_val/
      ...
    - kinetics_train_label.json
    - keintics_val_label.json
  - nturgbd_raw/
    - nturgb+d_skeletons/     # from `nturgbd_skeletons_s001_to_s017.zip`
      ...
    - nturgb+d_skeletons120/  # from `nturgbd_skeletons_s018_to_s032.zip`
      ...
    - NTU_RGBD_samples_with_missing_skeletons.txt
    - NTU_RGBD120_samples_with_missing_skeletons.txt

Generating Data

1. NTU RGB+D

   • cd data_gen
   • python3 ntu_gendata.py
   • python3 ntu120_gendata.py
   • Time estimate is ~ 3hrs to generate NTU 120 on a single core (feel free to parallelize the code :))

2. Kinetics

   • python3 kinetics_gendata.py
   • ~ 70 mins to generate Kinetics data

3. Generate the bone data with:

   • python gen_bone_data.py --dataset ntu
   • python gen_bone_data.py --dataset ntu120
   • python gen_bone_data.py --dataset kinetics

Pretrained Models

• Download pretrained models for producing the final results on NTU RGB+D 60, NTU RGB+D 120, Kinetics Skeleton 400: [Dropbox][GoogleDrive][WeiYun]

• Put the folder of pretrained models at repo root:

- MS-G3D/
  - pretrained-models/
  - main.py
  - ...

• Run bash eval_pretrained.sh

Training & Testing

• The general training template command:

python3 main.py
  --config <config file>
  --work-dir <place to keep things (weights, checkpoints, logs)>
  --device <GPU IDs to use>
  --half   # Mixed precision training with NVIDIA Apex (default O1) for GPUs ~11GB memory
  [--base-lr <base learning rate>]
  [--batch-size <batch size>]
  [--weight-decay <weight decay>]
  [--forward-batch-size <batch size during forward pass, useful if using only 1 GPU>]
  [--eval-start <which epoch to start evaluating the model>]

• The general testing template command:

python3 main.py
  --config <config file>
  --work-dir <place to keep things>
  --device <GPU IDs to use>
  --weights <path to model weights>
  [--test-batch-size <...>]

• Template for joint-bone two-stream fusion:

python3 ensemble.py
  --dataset <dataset to ensemble, e.g. ntu120/xsub>
  --joint-dir <work_dir of your test command for joint model>
  --bone-dir <work_dir of your test command for bone model>

• Use the corresponding config files from ./config to train/test different datasets

• Examples

  • Train on NTU 120 XSub Joint
    • Train with 1 GPU:
      • python3 main.py --config ./config/nturgbd120-cross-subject/train_joint.yaml
    • Train with 2 GPUs:
      • python3 main.py --config ./config/nturgbd120-cross-subject/train_joint.yaml --batch-size 32 --forward-batch-size 32 --device 0 1
  • Test on NTU 120 XSet Bone
    • python3 main.py --config ./config/nturgbd120-cross-setup/test_bone.yaml
  • Batch size 32 on 1 GPU (BS 16 per forward pass by accumulating gradients):
    • python3 main.py --config <...> --batch-size 32 --forward-batch-size 16 --device 0

• Resume training from checkpoint

python3 main.py
  ...  # Same params as before
  --start-epoch <0 indexed epoch>
  --weights <weights in work_dir>
  --checkpoint <checkpoint in work_dir>

Notes

• It's recommended to linearly scale up base LR with > 2 GPUs (https://arxiv.org/pdf/1706.02677.pdf, Section 2.1) to use 16 samples per worker during training; e.g.

  • 1 GPU: --base-lr 0.05 --device 0 --batch-size 32 --forward-batch-size 16
  • 2 GPUs: --base-lr 0.05 --device 0 1 --batch-size 32 --forward-batch-size 32
  • 4 GPUs: --base-lr 0.1 --device 0 1 2 3 --batch-size 64 --forward-batch-size 64

• Unfortunately, different PyTorch/CUDA versions & GPU setups can cause different levels of memory usage, and so you may experience out of memory (OOM) on some machines but not others

  • 1080Ti GPUs with --half and --amp-opt-level 1 (default) are relatively more stable

• If OOM occurs, try using Apex O2 by setting --amp-opt-level 2. However, note that

  • NVIDIA Apex does not yet support nn.DataParallel for O2
    • NVIDIA/apex#227 (comment)
    • This means you may need to train on a single GPU when using O2
  • It may also impact the stability of training and/or the final performance

• Default hyperparameters are stored in the config files; you can tune them & add extra training techniques to boost performance

• The best joint-bone fusion result may not come from the best single stream models; for example, we provided 3 pretrained models for NTU RGB+D 60 XSub joint stream where the best fusion performance comes from the slightly underperforming model (~89.3%) instead of the reported (~89.4%) and the slightly better retrained model (~89.6%).

Acknowledgements

This repo is based on

• 2s-AGCN
• ST-GCN

Thanks to the original authors for their work!

Citation

Please cite this work if you find it useful:

@inproceedings{liu2020disentangling,
  title={Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition},
  author={Liu, Ziyu and Zhang, Hongwen and Chen, Zhenghao and Wang, Zhiyong and Ouyang, Wanli},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={143--152},
  year={2020}
}

Contact

Please email kenziyuliu AT outlook.com for further questions