Remove torch.cuda check in setup.py

[eskjorg]

Let's say you want to run docker build on a machine without a GPU (for later use on a machine with GPU). In that case, only the CUDA_HOME is not None should be enough if CUDA is installed in the base image.

torch.cuda.is_available() on the other hand would evaluate to false. Or is there any important reason keep that check?

See hack at:
#167 (comment)

[fmassa]

Thanks for the PR, but I'm afraid that this will cause problems for a number of other users, so I'd rather not bias towards this use-case I think.

[eskjorg]

Thanks for the PR, but I'm afraid that this will cause problems for a number of other users, so I'd rather not bias towards this use-case I think.

Just out of curiosity, do you know any example use cases in which problems could arise?

[fmassa]

Say you have some bits of the CUDA installed in your machine (even if you don't have GPUs). That was a requirement for Caffe to install if I remember properly.
Now you'd try to compile the extensions and I'm really not sure this would work out fine or give weird problems to CPU-only users.

The docker setup is an optional setup in maskrcnn-benchmark. The recommended installation instructions are by using conda, and this has worked out well so far, so I'd rather not change it.

[miguelvr]

The PR in #612 should fix the issue.

[eskjorg]

Great! Closing this.