Segmentation fault encountered when entering the second epoch with num_workers>0 about torchdistill

[yoshitomo-matsubara]

@RulinShao

Recently, I faced a similar error with a different project on a new environment and resolved the issue by stopping thread control in evaluate.
Fetch and try the updated image_classification.py

[RulinShao]

Thanks a lot for your help!!! Stopping the thread control works for me. I'm closing this issue now.

[yoshitomo-matsubara]

Hi @RulinShao

I have never seen the error before and would need more detail.
Could you clarify

1. whether or not kd_main.py is identical to examples/image_classification.py I provided,
2. exact command you used to run kd_main.py and config file, and
3. environment info (OS, version of Python, torch, torchvision, torchdistill, etc)
   ?

Also, the following link could help you, and it may be an environmental issue
facebookresearch/detectron2#954

Thank you

[RulinShao]

Hi @yoshitomo-matsubara

Thanks for your reply! The kd_main.py is identical to examples/image_classification.py and I ran it by

python3 kd_main.py --config configs/ilsvrc2012/single_stage/kd/resnet18_from_resnet34.yaml --log log/ilsvrc2012/kd/resnet_from_vit.txt

where the config file is identical to torchdistill/configs/sample/ilsvrc2012/single_stage/kd/resnet18_from_resnet152.yaml except that I changed the teacher model from resnet152 to resnet34 for faster debugging speed.

Some details of the environment:
Python==3.7.9,
torch==1.8.1+cu102,
torchvision==0.9.1+cu102,
and torchdistill was just git cloned few days ago.

And thanks for the link which said his environment issue was solved by using sudo apt-get install package. However, as I use Amazon Linux, I need use yum install instead of apt-get install and these packages are not found by yum. I guess it's caused by the dependencies of some low-level libraries indeed where I have little knowledge. Kindly you could help me find out what is the problem. Currently I can only train one epoch each time and rerun the script while loading the ckpt when num_workers>0, or just set num_worker=0 which is quite time-consuming.

Thanks again for your help!

[yoshitomo-matsubara]

Hi @RulinShao

Thank you for the info.
I found this discussion useful for you. One of the users in this thread provides a solution for CentOS, which should be compatible with Amazon Linux as these are based on RHEL.

Also ICYMI, most of the config files under torchdistill/configs/sample are not tuned but used for debug as described torchdistill/configs/. If you want to see the improvements over standard training after debugging, you should either tune the hyperparameters in the config file or use some of those under torchdistill/configs/official/.

Hope this helps