Network Dissection of a DQN Agent Playing Super Mario Bros

(gif from gifer.com)

This repository contains three modules:

1. rl: Train a DQN agent in the gym environment that plays Super Mario Bros;
2. segmentation: Train a segmenter model that produces the semantic segmentations of Super Mario Bros game scenes (following Javier Montalvo et al.'s work);
3. dissect: Performs network dissection on the trained DQN network (following David Bau's work).

The first part is rather independent and is included as a git submodule. For each module, check their README.mds for detailed information. Note that the first module runs with Python 3.9.x and the other two run with Python 3.10.x. The third part should be run in UNIX environments for the use of symlinks.