GPU in FINN docker

[btwbtw01]

Hi,
I want to train VGG16 model using Cifar10 dataset. However, the docker has no GPU to train model, and use CPU to train model is very slow. How can I install GPU in the docker when I have GPU in mt computer?

[nhma20]

I also cannot get the GPU to work inside the docker container. Instead, I have a Brevitas/Pytorch environment outside the docker where I can do GPU accelerated training. When done, I simply import the trained network inside the FINN docker environment and continue the workflow there.
Not a solution to your issue, but maybe it's still useful to you.

[fpjentzsch]

Hi,
is NVIDIA Container Toolkit installed correctly?

I'm on the dev branch and use FINN from within WSL2. GPU support works fine inside the container, but only if I set FINN_DOCKER_GPU=1 manually before running ./run_docker.sh. The automatic detection (see

finn/run-docker.sh

Line 94 in 963aba1

: ${FINN_DOCKER_GPU="$(docker info | grep nvidia | wc -m)"}

) doesn't seem to work in my scenario. Maybe this also works for you?

[btwbtw01]

Thanks for your suggestion! Set FINN_DOCKER_GPU=1 doesn't seem to work in my scenario either. I will try @nhma20 solution. Use Brevitas/Pytorch environment outside the docker first, then save model, import the trained network inside the FINN docker.