6823f7b27dc99eeeb3ba7be57dc89d2e1908422a,ml_perf/reference_implementation.py,,rl_loop,#,360

Before Change

    else:
      // Run eval, validation & selfplay sequentially.
      model_win_rate = wait(evaluate_trained_model(state))
      wait(validate(state, holdout_glob))
      wait(selfplay(state))

    if model_win_rate >= FLAGS.gating_win_rate:

After Change

        // If we"re bootstrapping a checkpoint, evaluate the newly trained model
        // against the target.
        if FLAGS.bootstrap:
            target_model_path = os.path.join(fsdb.models_dir(), "target.pb")
            sgf_dir = os.path.join(
                fsdb.eval_dir(),
                "{}-vs-target".format(state.train_model_name))
            win_rate_vs_target = wait(evaluate_model(
                state.train_model_path, target_model_path, sgf_dir))
            if (win_rate_vs_target >= FLAGS.bootstrap_target_win_rate and
                prev_win_rate_vs_target > 0):
                break
            prev_win_rate_vs_target = win_rate_vs_target

Instances