run_mujoco.py 文件源码

def train(env_id, num_timesteps, seed):
    env=gym.make(env_id)
    env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
    set_global_seeds(seed)
    env.seed(seed)
    gym.logger.setLevel(logging.WARN)

    with tf.Session(config=tf.ConfigProto()):
        ob_dim = env.observation_space.shape[0]
        ac_dim = env.action_space.shape[0]
        with tf.variable_scope("vf"):
            vf = NeuralNetValueFunction(ob_dim, ac_dim)
        with tf.variable_scope("pi"):
            policy = GaussianMlpPolicy(ob_dim, ac_dim)

        learn(env, policy=policy, vf=vf,
            gamma=0.99, lam=0.97, timesteps_per_batch=2500,
            desired_kl=0.002,
            num_timesteps=num_timesteps, animate=False)

        env.close()