Max Schwarzer*, Ankesh Anand*, Rishab Goel, R Devon Hjelm, Aaron Courville, Philip Bachman
This repo provides code for implementing the SPR paper
- ๐ฆ Install -- Install relevant dependencies and the project
- ๐ง Usage -- Commands to run different experiments from the paper
To install the requirements, follow these steps:
# PyTorch
conda install pytorch torchvision -c pytorch
export LC_ALL=C.UTF-8
export LANG=C.UTF-8
# Install requirements
pip install -r requirements.txt
# Finally, clone the project
git clone https://github.com/mila-iqia/spr
The default branch for the latest and stable changes is release
.
- To run SPR with augmentation
python -m scripts.run --public --game pong --momentum-tau 1.
- To run SPR without augmentation
python -m scripts.run --public --game pong --augmentation none --target-augmentation 0 --momentum-tau 0.01 --dropout 0.5
When reporting scores, we average across 10 seeds.
.
โโโ scripts
โ โโโ run.py # The main runner script to launch jobs.
โโโ src
โ โโโ agent.py # Implements the Agent API for action selection
โ โโโ algos.py # Distributional RL loss
โ โโโ models.py # Network architecture and forward passes.
โ โโโ rlpyt_atari_env.py # Slightly modified Atari env from rlpyt
โ โโโ rlpyt_utils.py # Utility methods that we use to extend rlpyt's functionality
โ โโโ utils.py # Command line arguments and helper functions
โ
โโโ requirements.txt # Dependencies