Example #1
0
def test_ddpg(args=get_args()):
    env = gym.make(args.task)
    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    args.max_action = env.action_space.high[0]
    # train_envs = gym.make(args.task)
    train_envs = VectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)])
    # test_envs = gym.make(args.task)
    test_envs = SubprocVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.test_num)])
    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    train_envs.seed(args.seed)
    test_envs.seed(args.seed)
    # model
    actor = Actor(
        args.layer_num, args.state_shape, args.action_shape,
        args.max_action, args.device
    ).to(args.device)
    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
    critic = Critic(
        args.layer_num, args.state_shape, args.action_shape, args.device
    ).to(args.device)
    critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
    policy = DDPGPolicy(
        actor, actor_optim, critic, critic_optim,
        args.tau, args.gamma, GaussianNoise(sigma=args.exploration_noise),
        [env.action_space.low[0], env.action_space.high[0]],
        reward_normalization=True, ignore_done=True)
    # collector
    train_collector = Collector(
        policy, train_envs, ReplayBuffer(args.buffer_size))
    test_collector = Collector(policy, test_envs)
    # log
    writer = SummaryWriter(args.logdir + '/' + 'ddpg')

    def stop_fn(x):
        return x >= env.spec.reward_threshold

    # trainer
    result = offpolicy_trainer(
        policy, train_collector, test_collector, args.epoch,
        args.step_per_epoch, args.collect_per_step, args.test_num,
        args.batch_size, stop_fn=stop_fn, writer=writer, task=args.task)
    assert stop_fn(result['best_reward'])
    train_collector.close()
    test_collector.close()
    if __name__ == '__main__':
        pprint.pprint(result)
        # Let's watch its performance!
        env = gym.make(args.task)
        collector = Collector(policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
        collector.close()
def test_sac(args=get_args()):
    torch.set_num_threads(1)
    env = gym.make(args.task)
    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    args.max_action = env.action_space.high[0]
    # you can also use tianshou.env.SubprocVectorEnv
    # train_envs = gym.make(args.task)
    train_envs = SubprocVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)])
    # test_envs = gym.make(args.task)
    test_envs = SubprocVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.test_num)])
    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    train_envs.seed(args.seed)
    test_envs.seed(args.seed)
    # model
    actor = ActorProb(
        args.layer_num, args.state_shape, args.action_shape,
        args.max_action, args.device, unbounded=True
    ).to(args.device)
    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
    critic1 = Critic(
        args.layer_num, args.state_shape, args.action_shape, args.device
    ).to(args.device)
    critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
    critic2 = Critic(
        args.layer_num, args.state_shape, args.action_shape, args.device
    ).to(args.device)
    critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
    policy = SACPolicy(
        actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,
        args.tau, args.gamma, args.alpha,
        [env.action_space.low[0], env.action_space.high[0]],
        reward_normalization=True, ignore_done=True)
    # collector
    train_collector = Collector(
        policy, train_envs, ReplayBuffer(args.buffer_size))
    test_collector = Collector(policy, test_envs)
    # train_collector.collect(n_step=args.buffer_size)
    # log
    log_path = os.path.join(args.logdir, args.task, 'sac', args.run_id)
    writer = SummaryWriter(log_path)

    def stop_fn(x):
        return x >= env.spec.reward_threshold

    # trainer
    result = offpolicy_trainer(
        policy, train_collector, test_collector, args.epoch,
        args.step_per_epoch, args.collect_per_step, args.test_num,
        args.batch_size, stop_fn=stop_fn,
        writer=writer, log_interval=args.log_interval)
    assert stop_fn(result['best_reward'])
    train_collector.close()
    test_collector.close()
    if __name__ == '__main__':
        pprint.pprint(result)
        # Let's watch its performance!
        env = gym.make(args.task)
        collector = Collector(policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
        collector.close()
Example #3
0
def test_sac(args=get_args()):
    env = gym.make(args.task)
    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    args.max_action = env.action_space.high[0]
    # train_envs = gym.make(args.task)
    train_envs = VectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)])
    # test_envs = gym.make(args.task)
    test_envs = VectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.test_num)])
    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    train_envs.seed(args.seed)
    test_envs.seed(args.seed)
    # model
    actor = ActorProb(args.layer_num, args.state_shape, args.action_shape,
                      args.max_action, args.device).to(args.device)
    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
    critic1 = Critic(args.layer_num, args.state_shape, args.action_shape,
                     args.device).to(args.device)
    critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
    critic2 = Critic(args.layer_num, args.state_shape, args.action_shape,
                     args.device).to(args.device)
    critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)

    if args.auto_alpha:
        target_entropy = -np.prod(env.action_space.shape)
        log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
        alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
        alpha = (target_entropy, log_alpha, alpha_optim)
    else:
        alpha = args.alpha

    policy = SACPolicy(actor,
                       actor_optim,
                       critic1,
                       critic1_optim,
                       critic2,
                       critic2_optim,
                       args.tau,
                       args.gamma,
                       alpha,
                       [env.action_space.low[0], env.action_space.high[0]],
                       reward_normalization=args.rew_norm,
                       ignore_done=True,
                       exploration_noise=OUNoise(0.0, args.noise_std))
    # collector
    train_collector = Collector(policy, train_envs,
                                ReplayBuffer(args.buffer_size))
    test_collector = Collector(policy, test_envs)
    # train_collector.collect(n_step=args.buffer_size)
    # log
    log_path = os.path.join(args.logdir, args.task, 'sac')
    writer = SummaryWriter(log_path)

    def save_fn(policy):
        torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))

    def stop_fn(x):
        return x >= env.spec.reward_threshold

    # trainer
    result = offpolicy_trainer(policy,
                               train_collector,
                               test_collector,
                               args.epoch,
                               args.step_per_epoch,
                               args.collect_per_step,
                               args.test_num,
                               args.batch_size,
                               stop_fn=stop_fn,
                               save_fn=save_fn,
                               writer=writer)
    assert stop_fn(result['best_reward'])
    train_collector.close()
    test_collector.close()
    if __name__ == '__main__':
        pprint.pprint(result)
        # Let's watch its performance!
        env = gym.make(args.task)
        collector = Collector(policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
        collector.close()