vwxyzjn / gym-pysc2

Gym wrapper for pysc2

gym-pysc2

Gym wrapper for pysc2. This repo contains an implementation that honors the gym's API and supports the rendering function. So it is possible to use the gym.wrappers.Monitor class to create videos of agents playing the game.

This is still a work in progress but you can see some example runs here

https://wandb.ai/cleanrl/cleanrl.benchmark/reports/Gym-pysc2-Benchmark--VmlldzoyNTEyMTc

Installation

# First step: install SC2 
## type `iagreetotheeula` to unzip SC2
wget -O ~/sc2.zip http://blzdistsc2-a.akamaihd.net/Linux/SC2.4.10.zip && \
unzip ~/sc2.zip -d ~/ && \
rm ~/sc2.zip

## hack: fix the error message
## ffmpeg: ~/StarCraftII/Libs/libstdc++.so.6: version `GLIBCXX_3.4.26'
## but make sure you have ffmpeg installed
mv ~/StarCraftII/Libs/libstdc++.so.6 ~/StarCraftII/Libs/libstdc++.so.6.temp

# Second step: install pysc2 
git clone https://github.com/vwxyzjn/gym-pysc2
cd gym-pysc2
<!-- pip install gym_pysc2 --upgrade -->

# Optional: construct virtualenv via `pyenv`
pyenv install -s $(sed "s/\/envs.*//" .python-version)
pyenv virtualenv $(sed "s/\/envs\// /" .python-version)
pip install --upgrade pip wheel

# install dependencies for running the training script
pip install -r requirements.txt

Get started

To test installation, run the following code.

python ppo_sc2.py \
    --gym-id SC2MoveToBeacon-v0 \
    --num-envs 1 \
    --num-steps 256 \
    --cuda True \
    --capture-video

Then open another terminal to check tensorboard

tensorboard --logdir runs

You could also check the videos of agents playing the game in the videos folder

Run with wandb integration

# login wandb by using `wandb login`
for seed in {1..1}
do
    (sleep 0.3 && nohup xvfb-run -a python ppo_sc2.py \
    --total-timesteps 100000000 \
    --gym-id SC2MoveToBeacon-v0 \
    --num-envs 8 \
    --num-steps 256 \
    --wandb-project-name gym_pysc2 \
    --prod-mode \
    --cuda True \
    --capture-video \
    --seed $seed) >& /dev/null &
done