vision-dqn

Get started

Prerequisites:

• Python 3.8+
• Poetry

To run experiments locally, give the following a try:

poetry install

To run our ResNet-based model on <game> [Breakout, SpaceInvaders, MsPacman]:

poetry run python dqn_atari.py --exp-name=<game>_resnet --encoder resnet --gym-id <game>NoFrameskip-v4

To run DCGAN-based model on <game> [Breakout, SpaceInvaders, MsPacman] (after pretraining DCGAN model):

poetry run python dqn_atari.py --exp-name=<game>_dcgan --encoder dcgan --dcgan_path <dcgan_path> --gym-id <game>NoFrameskip-v4

Pretraining DCGAN

Open the dcgan-atari-train.ipynb on Google Colab with GPU runtime and run all blocks. It should save checkpoints in the temporary folder on Colab, from which the Discriminator checkpoint path must be assigned as <dcgan_path>.