Add Dockerfile

[Max-We]

trafficstars

I created a simple Dockerfile that installs all required python & system dependencies and runs one of the python scripts (currently run_baseline_parallel.py, but this can easily be adapted). This pull request is related to #111

In case we want to support multiple commands (not just run_baseline_parallel.py) it would be appropriate to create multiple Dockerfiles or an entry point script.

[PWhiddy]

Hi! You've tested this dockerfile, does it need things like ffmpeg for example? Also I wonder if starting a container to run a single command makes sense, does this create a good workflow for running experiments that doesn't result in deleting results after each run or having many stale containers sitting around? Just thinking through these things ...

[Max-We]

You've tested this dockerfile, does it need things like ffmpeg for example?

For the installation (pip), it does not, however I haven't tried training an entire model with it.

Also I wonder if starting a container to run a single command makes sense

What workflows do you have in mind?

[joshhsoj1902]

I was struggling getting this to run under python 3.12. so I decided to give this docker image a go. I made a few small tweaks which don't feel worth contributing but are maybe worth noting here:

Dockerfile

FROM python:3.10-bookworm

# Set the working directory in docker
WORKDIR /pokemon

# Install system dependencies
RUN apt-get update && apt-get install -y \
  build-essential \
  git && \
  # Clean up apt cache to reduce image size.
  apt-get clean && \
  rm -rf /var/lib/apt/lists/*

RUN apt-get update && apt-get install -y \
  ffmpeg && \
  # Clean up apt cache to reduce image size.
  apt-get clean && \
  rm -rf /var/lib/apt/lists/*

RUN mkdir baselines

COPY baselines/requirements.txt baselines

# Install Python dependencies
RUN pip install --no-cache-dir -r baselines/requirements.txt

RUN pip install websockets

# Copy source-code
# COPY . .

WORKDIR /pokemon/baselines

# Command to run at container start
CMD [ "python", "run_baseline_parallel_fast.py" ]

The main differences are:

• Changed the final workdir so it can find the rom
• Don't copy all the source into the image, instead mount the local dir at runtime

I then run it using this docker-compose

docker-compose.yaml

version: "3.4"
services:
  pokemonredexperiments:
    image: joshhsoj1902/pokemonredexperiments:latest
    volumes:
      - ./:/pokemon
    restart: always

And for complete clarity here are the build & run commands:

    docker build --progress=plain -t joshhsoj1902/pokemonredexperiments:latest .

	docker compose up

---

I don't know if this is completely working or not since I wasn't able to get it to run outside of docker, but I was able to get things to show up in https://pwhiddy.github.io/pokerl-map-viz/. in terms of the session screenshots, The "curframe" screenshots work fine, but the reward frames screenshots are usually black/white boxes (with a few exceptions, so maybe it's actually working)

I look forward to playing with this more!

---

Edit: I just updated the dockerfile to install ffmpeg and the video generation seems to work (for anyone just getting started it's off by default, look for the save_video setting in the env setup). Sometimes it crashes on startup but usually it runs fine (I've only really ran it for about 30 min at this point, but if I don't update this post now I'll forget)