In maskma.py, the network architecture of MaskMA, as well as the logic for training and testing, are implemented. The "forward" function serves as the main entry point, where the "if not eval" branch represents the training logic, and the "else" branch represents the evaluation logic.

We provide a collection of replays showcasing MaskMA's performance on various maps in SMAC. The 'replay/pretrain' directory contains replays of MaskMA on 11 pretraining maps, while the 'replay/downstream' directory includes replays of MaskMA in three downstream tasks: varied policies collaboration, teammate malfunction, and ad hoc team play. 