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.
- 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)
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.
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)
instead of
python ./data_gen/ntu_gen_preprocess.py
you should
cd datagen, python ntu_gen_preprocess.py, cd ..
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.