docker file gpu added

[muaz-urwa]

No description provided.

[andrewjong]

Nice!! Many thanks @muaz-urwa. I'll verify it on my machine later today, and update the README to mention this.

[muaz-urwa]

Sure. Let me know if you face any issues.
Few pointers:

1. This is the docker command I used:
   docker run -it -v ${PWD}/checkpoints:/app/SwapNet/checkpoints -v ${PWD}/data:/app/SwapNet/data -v ${PWD}/results:/app/SwapNet/results --gpus all swapnet

So I am mounting the data, model and results folders.

2. After running the docker:
   a) cd into Swapnet
   b) source activate swapnet

3. Inference command:
   python inference.py --checkpoint checkpoints/deep_fashion
   --dataroot data/deep_fashion
   --shuffle_data True

4. If you face issues with roialign, and the error says that it was not compiled with GPU support. Then GPU was not available while building the image. In this case few checks at the end of dockerfile (last three commands) will have incorrect output. See comments in docker file. If this happens set your default docker runtime to nvida following instructions here.

[andrewjong]

For me it crashes on /bin/sh: 1: nvidia-smi: not found. I think nvidia-smi may not be able to run during the image build; do you have knowledge on this @muaz-urwa ? Perhaps we can move checking nvidia-smi to post-installation instructions.

[andrewjong]

I updated this PR to merge into this repo's feat/docker branch, as I want to make some edits. Thanks for your hard work Urwa, this will be merged into master soon.

[muaz-urwa]

Okay. In that case try this:
Only way to have gpu acess during build is to set default docker runtime as nvidia. We need gpu access during build to compile roi library with gpu support.

Find the file: /etc/docker/daemon.json
and add "default-runtime": "nvidia" as shown below.

my /etc/docker/daemon.json
{
"runtimes": {
"nvidia": {
"path": "nvidia-container-runtime",
"runtimeArgs": []
}
},
"default-runtime": "nvidia"
}

[andrewjong]

Alternatively, I can host this docker image on Dockerhub. Then people don't need a GPU to build and can just pull the image.

Great suggestion for building with GPU though.

[muaz-urwa]

Also you will need to install nvidia-docker to acess gpu devices from containers. Here is their repo : link

I kept these checks in docker file to make sure that gpus were accesible during build, otherwise roi align would crash later on.

Yes, hosting seems to be a good idea.