PySC2 bot in Pytorch

Reference:

Relies on the repo https://github.com/simonmeister/pysc2-rl-agents for implementations of action / observation space pre-processing, network architectures.

Entire backend is in Pytorch, with TF for tensorboard.

To run training:

# Overwrite existing model
python run.py --experiment_id my_experiment --envs 32 --map MoveToBeacon --overwrite
# Load existing model
python run.py --experiment_id my_experiment --envs 32 --map MoveToBeacon
# With multiple GPUs
CUDA_VISIBLE_DEVICES=1,2 python run_sv.py --data_dir data/

Result:

MoveToBeacon	CollectMineralShards

FindAndDefeatZerglings

TODO:

• Train on other mini-games
• Use replay data
• Optimize the Runner to work with Torch tensor instead of numpy array
• Multi-GPU training