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

Before Change

    state.iter_num += 1
    wait(train(state, tf_records))

    if FLAGS.parallel_post_train:
      // Run eval, validation & selfplay in parallel.
      model_win_rate, _, _ = wait([
          evaluate_trained_model(state),
          validate(state, holdout_glob),
          selfplay(state)])
    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:
      // Promote the trained model to the best model and increment the generation
      // number.
      state.best_model_name = state.train_model_name

After Change

                "{}-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

        if model_win_rate >= FLAGS.gating_win_rate:
            // Promote the trained model to the best model and increment the
            // generation number.

Instances