Pytorch_geometric: Memory usage of graclus pooling

[jacksoncsy]

Hi Matthias,

Thanks for this great project! However, I have got an issue with graclus pooling. GPU memory increases after each pooling operation, while this part of memory is seldom released. As I have a large dataset, the training had to stop quite early because of insufficient memory.
I am not sure if there is memory leak in graclus.py? I am looking into this part of code now. Meanwhile, would you mind checking this as well?

Many thanks,
Shiyang

[rusty1s]

Hello Shiyang,
thanks for reporting this issue. I will look into it.

[rusty1s]

I've updated torch-cluster (version 1.1.2). There was indeed a memory leak. Can you confirm that it's now working?

[chrsmrrs]

No, issue remains.

Chris

Am 17.05.2018 um 17:40 schrieb Matthias Fey:
>

I've updated |torch-cluster| (version 1.1.2). There was indeed a
memory leak. Can you confirm that it's now working?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/rusty1s/pytorch_geometric/issues/5#issuecomment-389911816,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AYJkdFerLTcU3b4qxF0bQKeCpfFfFVZLks5tzZnvgaJpZM4UBSI7.

--
Christopher Morris
Algorithm Engineering,
Department of Computer Science,
TU Dortmund University

https://ls11-www.cs.tu-dortmund.de/staff/morris

[jacksoncsy]

Brilliant! I think this issue is solved. I can see that memory is fixed after each epoch (2 Mb to be exact).

BTW, may I ask where exactly this memory leak happens? I tried to locate the problem yesterday, unfortunately my knowledge on CUDA is quite limited :(

Many thanks,
Shiyang

[rusty1s]

https://github.com/rusty1s/pytorch_cluster/blob/master/aten/THC/THCColor.cuh#L33
I didn't free the bernoulli tensor.