search

neptune-ai / open-solution-mapping-challenge

Open solution to the Mapping Challenge :earth_americas:
https://www.crowdai.org/challenges/mapping-challenge
MIT License
380 stars 96 forks source link

How to use tensor cores in V100 GPU? #188

Closed carbonox-infernox closed 6 years ago

carbonox-infernox commented 6 years ago

I've been reading into using tensor cores for training, and it would appear that pytorch is capable of using them as of release 17.09.

This question is directed toward the devs: Would it be possible for me to get open-solution-mapping-challenge to use tensor cores by reinstalling various things from source, such as pytorch and torchvision? As suggested here: https://discuss.pytorch.org/t/pytorch-on-v100-gpu/9331

Or would I need to go into this project's source code in order to change the way it handles things to make it tensor-core compatible?

kamil-kaczmarek commented 6 years ago

Hi @carbonox-infernox,

We did not test our solution on V100 GPU, but feel free to try it!

Best, Kamil

carbonox-infernox commented 6 years ago

@kamil-kaczmarek I would like to, but I was hoping for suggestions or maybe some insight into whether the code base would handle the FP16 data and operations. Could you reopen this please? I don't think anyone will see this if it's closed. Besides, it's an interesting potential feature and I think others might be interested in this.

kamil-kaczmarek commented 6 years ago

Hey @carbonox-infernox,

As I mentioned before we did not test V100 GPUs, and this was outside the scope of our project. I think, however, that it might be good excercise for you to try to configure/ setup this repo to use them efficiently.

This is also the reason why made this code open source :point_right: everybody can take it and try different things with it :)

Best of luck with your project!

Best, Kamil