Skip to content

Latest commit

 

History

History
79 lines (43 loc) · 2.04 KB

Notes for Reproduction.md

File metadata and controls

79 lines (43 loc) · 2.04 KB

Notes for Reproducing AS-GCN

Here I wrote down some of the issues/errors you might come across when reproducing AS-GCN and my ways of getting around. see the AS-GCN Paper as well as their Github. Hope it helps.

Environment

  • python=3.6
  • CUDA 10 (for NSCC) or 8 (for lab GPU server)
  • pytorch & torchvision (corresponding version with CUDA)
  • pyyaml, argparse, tqdm(not mentioned but used), h5py (needed for torchlight), matplotlib (not mentioned but used)

1. torchlight setup

Their Github gives following command:

cd torchlight, python setup.py, cd..

Clearly it should be:

cd torchlight, python setup.py install, cd..

refer to the code of ST-GCN.

Besides, the torchlight setup.py from AS-GCN's repo seems not to work properly with python 3.6. I use the torchlight from ST-GCN instead. And it works well.

2. Dataset unzipping

in .data_gen/ntu_gen_preprocess.py, line 143, instead of

gendata(arg.data_path, out_path, arg.ignored_sample_path, benchmark=b, part=sn)

it should be

gendata(arg.data_path, out_path, arg.ignored_sample_path, benchmark=b, set_name=sn)

3. Data preprocessing

instead of

python ./data_gen/ntu_gen_preprocess.py

you should

cd datagen, python ntu_gen_preprocess.py, cd ..

4. Train AIM

When running on servers and using PBS job scheduler, you might wish to add -koed to your pbs file, to indicate that you wish to have real-time streaming of your err & out file.

in main.py, comment out line 13, which is following:

processors['demo'] = import_class('processor.demo.Demo')

Another obvious typo from their Github: when training on cross subject, load configurations from the directory ntu-xsub; when training on cross view, load configurations from the directory nut-xview.