RocAlphaGo.data

Version controlled data for the Rochester-NRT/AlphaGo repository.

Where is the data?

Still working on it. Follow progress in the main repository. There is enough here to load and run a few partially-trained policy networks as standalone players.

Directory structure

scripts/ contains short code snippets that we've found useful

expert_play/ contains datasets of SGF files. We do not have permission to release all datasets that we are using, hence some models are trained on datasets not present in this directory

models/ contains json model specifications. For example

• models/CNNPolicy/12plane_default.json specifies a CNNPolicy object with the features ['board', 'ones', 'turns_since']
• models/CNNPolicy/20plane_default.json is likewise, with the features ['board', 'ones', 'turns_since', 'liberties']
• models/CNNPolicy/46plane_default.json is likewise with all features except ladders

training_results/ contains trained models and epoch checkpoints.