RuntimeError: CUDA out of memory

[happygirlzt]

Hi there, thank you very much for open-sourcing the work!
I wonder what devices you used for the work. Since I tried to run the training in a machine with 8 Tesla V100-SXM2-16GB, but cannot make it. Besides, I found the code would only utilize 2 GPUs, although I did not specify. I modified the device setting inside run.py, but still cannot change the fact that only 2 GPUs are used.
Please kindly suggest. Thank you in advance!

[happygirlzt]

Hi @pkuzqh , I've got another issue when running the code.

[pkuzqh]

If you want to change the batch size, you need to change the number in the dict "args". If you want to use multiple GPUs, you need to modify "model = nn.DataParallel(model, device_ids=[0, 1])".

[happygirlzt]

Hi @pkuzqh, thank you for the reply. The cuda out of memory issue has been resolved. However, I found the new error above. Please kindly suggest, thanks.

[pkuzqh]

How many GPUs do you use? And the batch size?

[happygirlzt]

3, I indicated in the train() that: device_ids=[1,2,3]
the batch size is 16

[pkuzqh]

You need to change the number "4" in line 103-106 to a multiple of 3. And the batch size needs to be a multiple of 3.

[happygirlzt]

OK, thank you very much @pkuzqh ! It can now run. However, I saw in the train(), the number of epochs is 100000, for epoch in range(100000): is that true?

[happygirlzt]

BTW, for inference, it looks like the testDefect4j.py can only use 1 GPU? Since I have 4 GPUs, only one was used, and it caused an OOM issue.

[pkuzqh]

you can use "nn.DataParallel" to use multiple gpus in testDefect4J.py