run_dqn_atari.py 文件源码

def main():
    arguments = docopt.docopt(__doc__)

    # Run training
    seed = 0  # Use a seed of zero (you may want to randomize the seed!)
    env = get_env(arguments['--envid'], seed)
    with get_session() as session:

        model = arguments['--model'].lower()
        num_filters = int(arguments['--num-filters'])
        batch_size = int(arguments['--batch-size'])
        print(' * [INFO] %s model (Filters: %d, Batch Size: %d)' % (
            model, num_filters, batch_size))

        save_path = atari_learn(
            env,
            session,
            num_timesteps=int(arguments['--timesteps']),
            num_filters=num_filters,
            model=model,
            batch_size=batch_size,
            restore=arguments['--restore'],
            checkpoint_dir=arguments['--ckpt-dir'],
            learning_starts=arguments['--learning-starts'])
        reader = tf.train.NewCheckpointReader(save_path)
        W = reader.get_tensor('q_func/action_value/fully_connected/weights')
        print('Largest entry:', np.linalg.norm(W, ord=np.inf))
        print('Frobenius norm:', np.linalg.norm(W, ord='fro'))