Skip to content
/ UGCN Public

Unsupervised Skeleton-based action recognition network

Notifications You must be signed in to change notification settings

eoeair/UGCN

Repository files navigation

UGCN

这基本是一个稳定版本,足够简练且平稳,目前在高性能优化

数据处理

  1. Get skeleton of each performer:python get_raw_skes_data.py
  2. Remove the bad skeleton:python get_raw_denoised_data.py
  3. Transform the skeleton to the center of the first frame:python seq_transformation.py

Directory Structure

- data/
  - ntu/
  - ntu120/
  - nturgbd_raw/
    - nturgb+d_skeletons/     # from `nturgbd_skeletons_s001_to_s017.zip`
      ...
    - nturgb+d_skeletons120/  # from `nturgbd_skeletons_s018_to_s032.zip`
      ...

模型训练

!OMP_NUM_THREADS=4 torchrun --standalone --nproc_per_node=1 --nnodes=1 pretrain.py --config config/train_cs.yaml --checkpoint-dir runs/pretrain/cs

注意:

  1. 默认启用了checkpoint机制,注意checkpoint只支持单阶段,例如预训练得到的权重只服务于预训练阶段,如果要将权重用到测试阶段,则必须使用gcn.pth这样的权重
  2. 默认将生成文件统一放入runs/pretrain

模型测试

python evaluate.py --epochs 100 --dev cuda:0 --pretrained runs/pretrain/cs/gcn.pth --config config/evaluate_cs.yaml --checkpoint-dir runs/evaluate/cs/

依赖

torch pyyaml tensorboardX einops h5py scikit-learn tqdm

About

Unsupervised Skeleton-based action recognition network

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published