VizDoom with Jupyter setup in Docker

Up and running

To get going as fast as possible (without visuals) you can do this. If you're training on servers - this approach might be the only resonable thing to do.

with CPU:

docker build --no-cache -t doom .

docker run -d -v $(pwd):/project -p 8888:8888 -p 6006:6006 doom jupyter lab --ip 0.0.0.0 --allow-root --no-browser --NotebookApp.token='' && open http:localhost:6006 && open http:localhost:8888

with GPU (assumes you're setup for gpu - if not see the section 'Up and running with GPU' below)

docker build -f ./Dockerfile.gpu --no-cache -t doom-gpu .

docker run --runtime=nvidia -d -v $(pwd):/project -p 8888:8888 -p 6006:6006 doom-gpu jupyter lab --ip 0.0.0.0 --allow-root --no-browser --NotebookApp.token='' && xdg-open http:localhost:6006 && xdg-open http:localhost:8888

Up and running with visuals

It is actually possible to see the frame by frame through docker even though it takes a bit of setup:

This works with the following versions on a host machine running Mac OS 10.14.3

• XQuartz 2.7.11 (xorg-server 1.18.4)
• Docker engine 18.09.2
• socat 1.7.3.2

0. Install XQuartz and socat

$ brew install socat
$ brew cask reinstall xquartz

After installation - restart your computer

1. Close any 6000 On a new terminal, verify that nothing is running on port 6000:

$ lsof -i TCP:6000
$

If there is anything, kill the process

2. Listen on port 6000 Open a socket on that port and keep the terminal open

$ socat TCP-LISTEN:6000,reuseaddr,fork UNIX-CLIENT:\"$DISPLAY\"

3. Verify 6000 is open In a new terminal, verify if it is opened

$ lsof -i TCP:6000
COMMAND   PID     USER   FD   TYPE             DEVICE SIZE/OFF NODE NAME
socat   29298 mdesales    5u  IPv4 0xe21e43ca9d99bf1d      0t0  TCP *:6000 (LISTEN)

4. Get going w. docker

docker build --no-cache -t doom .

docker run -d -v $(pwd):/project -e DISPLAY=docker.for.mac.host.internal:0 -p 8888:8888 -p 6006:6006 <image-name> jupyter lab --ip 0.0.0.0 --allow-root --no-browser --NotebookApp.token='' && open http:localhost:6006 && open http:localhost:8888

Up and running with GPU

Doing Deep RL using CPUs is very time consuming using GPUs is generally preferable. We train using quite a few different algorithms. As there are currently no Macbooks with nvidia gpu(the only feasible gpus for deep learning at the time of writing) we we only use mac for the preliminary testing. The actual training is done on ubuntu-machines. Here's what we need to do to get up and running docker:

1. Install GPU-driver. It's important to install a driver version that is compatible with the GPU arcitechture of your graphics card. If you don't you might get into trouble further down the road.
2. Install CUDA Toolkit
3. Install CUDNN
4. Install docker and nvidia-docker2 Setting up unfortunately isn't trivial. Thankfully, Pyimagesearch has a guide for doing steps 1-3. Step 4 is described here