Rewrite of https://github.com/vladfi1/phillip.
It is recommended that you first create a virtual env:
virtualenv venv
source venv/bin/activate
To install:
pip install -e .
First download the raw .slp files into the replays/ folder. A script is provided for the "Gang Steals" tournament.
./scripts/download-gang-steal.sh
Next, create the training data:
python dsmash/slippi/data.py --compress replays/Gang-Steals
This may take a while depending on how many replays you are processing. It will create a file called il-data/Gang-Steals_compressed.pkl. We can now train on this data:
python dsmash/imitation/train.py --data_path il-data/Gang-Steals_compressed.pkl
You will see mean_action_logp periodically printed, which measures how good your model is at predicting the actions taken by the recorded players. Higher is better, with the limit being zero.
Not implemented yet. It was implemented on the (now out of date) rllib code-path. Would be very nice to eval against a level 9 cpu while training.