Train TAO models to recognize actions in video.
Recent revisions are kept. Restoring one creates a new revision.
