Keras: How to specify which GPU to use on a multi-GPU machine?

[fish128]

I have a computer with 4 GPUs, and I want to train a few models at the same time on different GPUs. Is there a way to assign different GPUs when training different models?

[fish128]

ps: I am using TensorFlow as the backend.

[denlecoeuche]

You should look at #3333

[fish128]

Thanks @denlecoeuche. Seeting the environment variable CUDA_VISIBLE_DEVICES seems to be the easiest solution for my problem.

[sumsuddin]

in python add this, os.environ["CUDA_VISIBLE_DEVICES"]="1"

[sigmoidx]

Can two keras models run simultaneously with os.environ["CUDA_VISIBLE_DEVICES"]="0" and os.environ["CUDA_VISIBLE_DEVICES"]="1" settings on them?

[mingzhao03]

Can two keras models run simultaneously with os.environ["CUDA_VISIBLE_DEVICES"]="0" and os.environ["CUDA_VISIBLE_DEVICES"]="1" settings on them?

I wasn't sure whether it would work when I saw this comment since it was posed as a question, so I figured it might be useful for future visitors that yes, two keras models can run simultaneously with setting different os.environ["CUDA_VISIBLE_DEVICES"]. If you use tf.device instead, both instances will recognize all gpus, which would result in errors.