Closed A0nameless0man closed 2 years ago
ako_vs_ako is inefficient, because the emulator is designed to run 4 games in a batch, and akochan does not scale. If the challenger and the champion are the same (say akochan), with the same seed being reused 4 times, it will produce 4 identical game logs at a time by spawning 16 akochan processes in parallel, wasting 3x CPU time.
Fix train_grp may stuck for cuda device Add train data to git ignore Impl ako vs ako for easier data generation