Cuda not working in docker container

[javilinos]

Hi @omichel, I am trying to load a model composed by a CNN->LSTM->Policy trained using PPO. However, even If I set the parameter device=th.device('cpu'), it is not working as it seems that the feature extractor (cnn) constructor is ignoring this parameter in the SB3 implementation. I tried to install cuda in the container by installing pytorch which installs cuda as a dependency. Any tips on how I can configure the dockerfile so that the container succesfully configure cuda for this purpose if this is possible would be very appretiated. I will try to debug the issue meanwhile. Thanks.

[javilinos]

Solved. Now I have SB3 bug which I'll solve myself. Sorry for the actions mess, new to Docker.

[omichel]

Nice to read. 😃

[omichel]

It seems your Dockerfile could be made faster by building it FROM cyberbotics/webots.cloud:R2023a-ubuntu20.04-numpy which is actually defined here, and inherits from the main webots Dockerfile which in turn is created FROM nvidia/cuda:11.8.0-base-ubuntu22.04 base image. But maybe you actually need version 12 instead of version 11.8.0?

[javilinos]

Thanks for the tip, for now it is cached but when I change my Dockerfiile (which I will because I'm installing different versions of SB3 and SB3-contriib in the container than my system) I will change the image too. I became a little crazy trying to get it to work but at the end I was just missing these sentences:

• ENV NVIDIA_VISIBLE_DEVICES all
• ENV NVIDIA_DRIVER_CAPABILITIES compute,utility

Thanks!

[omichel]

If you don't mind we will add this to the main Webots Dockerfile, so that you won't have to bother any more about this problem in the future. See my PR here.

[javilinos]

Of course, go ahead!