lstm+ppo/sac

[1900360]

I wonder if lstm+ppo/sac could use in Tianshou? Since there are some problems.

[nuance1979]

There is no inherent reason that prevents a LSTM+PPO/SAC solution to work well in Tianshou and I have tried to fix it. However, I got distracted and have not been able to resume the work.

[1900360]

Hi @nuance1979!

import tianshou, gym, torch, numpy, sys
print(tianshou.__version__, gym.__version__, torch.__version__, numpy.__version__, sys.version, sys.platform)
0.4.9 0.19.0 1.11.0+cpu 1.22.4 3.9.13 (main, Aug 25 2022, 23:51:50) [MSC v.1916 64 bit (AMD64)] win32

I use sac + LSTM to play Pendulum-v0, but the target value has not been reached, and the reward has not increased. Is there any problem with my code?

My code is as follows:

import argparse
import os

import numpy as np
import pytest
import gym
import torch
from torch.utils.tensorboard import SummaryWriter

from tianshou.data import Collector, VectorReplayBuffer
from tianshou.policy import ImitationPolicy, SACPolicy
from tianshou.trainer import offpolicy_trainer
from tianshou.utils import TensorboardLogger
from tianshou.utils.net.common import Net
from tianshou.utils.net.continuous import Actor, RecurrentActorProb, RecurrentCritic, ActorProb, Critic
from tianshou.env import DummyVectorEnv, SubprocVectorEnv
from datetime import datetime

try:
    import envpool
except ImportError:
    envpool = None

def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--task', type=str, default='Pendulum-v0')
    parser.add_argument('--reward-threshold', type=float, default=None)
    parser.add_argument('--seed', type=int, default=1)
    parser.add_argument('--buffer-size', type=int, default=20000)
    parser.add_argument('--actor-lr', type=float, default=1e-4)
    parser.add_argument('--critic-lr', type=float, default=1e-4)
    parser.add_argument('--il-lr', type=float, default=1e-4)
    parser.add_argument('--gamma', type=float, default=0.99)
    parser.add_argument('--tau', type=float, default=0.005)
    parser.add_argument('--alpha', type=float, default=0.2)
    parser.add_argument('--auto-alpha', type=int, default=1)
    parser.add_argument('--alpha-lr', type=float, default=3e-4)
    parser.add_argument('--epoch', type=int, default=100000)
    parser.add_argument('--step-per-epoch', type=int, default=1000)
    parser.add_argument('--il-step-per-epoch', type=int, default=200)
    parser.add_argument('--step-per-collect', type=int, default=200)
    parser.add_argument('--update-per-step', type=float, default=0.1)
    parser.add_argument('--batch-size', type=int, default=128)
    parser.add_argument('--hidden-sizes', type=int, nargs='*', default=[128, 128])
    parser.add_argument(
        '--imitation-hidden-sizes', type=int, nargs='*', default=[128, 128]
    )
    parser.add_argument('--training-num', type=int, default=10)
    parser.add_argument('--test-num', type=int, default=1)
    parser.add_argument('--logdir', type=str, default='log')
    parser.add_argument('--render', type=float, default=0.)
    parser.add_argument('--rew-norm', action="store_true", default=False)
    parser.add_argument('--n-step', type=int, default=3)
    parser.add_argument(
        '--device', type=str, default='cuda' if torch.cuda.is_available() else 'cpu'
    )
    args = parser.parse_known_args()[0]
    return args

def sac_with_il(args):
    # if you want to use python vector env, please refer to other test scripts
    start_datetime = datetime.now()
    env = gym.make(args.task)
    train_envs = DummyVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)]
    )
    # test_envs = DummyVectorEnv(
    #     [lambda: gym.make(args.task) for _ in range(args.test_num)]
    # )
    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]
    if args.reward_threshold is None:
        default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250}
        args.reward_threshold = default_reward_threshold.get(
            args.task, env.spec.reward_threshold
        )
    # you can also use tianshou.env.SubprocVectorEnv
    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    # model
    layer_num = 4
    actor = RecurrentActorProb(
        layer_num=layer_num,
        state_shape=args.state_shape,
        action_shape=args.action_shape,
        max_action=args.max_action,
        device=args.device,
        unbounded=True
    ).to(args.device)
    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
    hidden_layer_size = 128
    critic1 = RecurrentCritic(
        layer_num=layer_num,
        state_shape=args.state_shape,
        action_shape=args.action_shape,
        hidden_layer_size=hidden_layer_size,
        device=args.device).to(args.device)
    critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)

    critic2 = RecurrentCritic(
        layer_num=layer_num,
        state_shape=args.state_shape,
        action_shape=args.action_shape,
        hidden_layer_size=hidden_layer_size,
        device=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)
        args.alpha = (target_entropy, log_alpha, alpha_optim)

    policy = SACPolicy(
        actor,
        actor_optim,
        critic1,
        critic1_optim,
        critic2,
        critic2_optim,
        tau=args.tau,
        gamma=args.gamma,
        alpha=args.alpha,
        reward_normalization=args.rew_norm,
        estimation_step=args.n_step,
        action_space=env.action_space
    )
    # collector
    train_collector = Collector(
        policy,
        train_envs,
        VectorReplayBuffer(args.buffer_size, len(train_envs), stack_num=4),
        exploration_noise=True
    )
    test_collector = None
    # train_collector.collect(n_step=args.buffer_size)
    # log
    log_path = os.path.join(args.logdir, args.task, 'sac', start_datetime.strftime('%Y%m%d/%H%M%S'))
    writer = SummaryWriter(log_path)
    logger = TensorboardLogger(writer)

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

    def stop_fn(mean_rewards):
        return mean_rewards >= args.reward_threshold

    # trainer
    train_collector.collect(n_step=20000, random=True)
    result = offpolicy_trainer(
        policy,
        train_collector,
        test_collector,
        args.epoch,
        args.step_per_epoch,
        args.step_per_collect,
        args.test_num,
        args.batch_size,
        update_per_step=args.update_per_step,
        stop_fn=stop_fn,
        save_best_fn=save_best_fn,
        logger=logger
    )

if __name__ == '__main__':
    args1 = get_args()
    sac_with_il(args1)

[1900360]

Hi @nuance1979! I coupled ppo and lstm together, but the training fails to converge, is my setup correct?

import argparse
import os
import pprint
import numpy as np
import pytest
import gym
import torch
from torch.distributions import Independent, Normal
from torch.utils.tensorboard import SummaryWriter

from tianshou.data import Collector, VectorReplayBuffer,ReplayBuffer
from tianshou.env import DummyVectorEnv
from tianshou.policy import PPOPolicy
from tianshou.trainer import OnpolicyTrainer
from tianshou.utils import TensorboardLogger
from tianshou.utils.net.common import ActorCritic, Net
from tianshou.utils.net.continuous import ActorProb, Critic,RecurrentActorProb,RecurrentCritic

def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--task', type=str, default='Pendulum-v0')
    parser.add_argument('--reward-threshold', type=float, default=None)
    parser.add_argument('--seed', type=int, default=1)
    parser.add_argument('--buffer-size', type=int, default=20000)
    parser.add_argument('--lr', type=float, default=5e-4)
    parser.add_argument('--gamma', type=float, default=0.95)
    parser.add_argument('--epoch', type=int, default=100)
    parser.add_argument('--step-per-epoch', type=int, default=200000)
    parser.add_argument('--episode-per-collect', type=int, default=20)
    parser.add_argument('--repeat-per-collect', type=int, default=2)
    parser.add_argument('--batch-size', type=int, default=256)
    parser.add_argument('--hidden-sizes', type=int, nargs='*', default=[512, 256])
    parser.add_argument('--training-num', type=int, default=20)
    parser.add_argument('--test-num', type=int, default=100)
    parser.add_argument('--logdir', type=str, default='log')
    parser.add_argument('--render', type=float, default=0.)
    parser.add_argument(
        '--device', type=str, default='cuda' if torch.cuda.is_available() else 'cpu'
    )
    # ppo special
    parser.add_argument('--vf-coef', type=float, default=0.25)
    parser.add_argument('--ent-coef', type=float, default=0.0)
    parser.add_argument('--eps-clip', type=float, default=0.2)
    parser.add_argument('--max-grad-norm', type=float, default=0.5)
    parser.add_argument('--gae-lambda', type=float, default=0.95)
    parser.add_argument('--rew-norm', type=int, default=1)
    parser.add_argument('--dual-clip', type=float, default=None)
    parser.add_argument('--value-clip', type=int, default=1)
    parser.add_argument('--norm-adv', type=int, default=1)
    parser.add_argument('--recompute-adv', type=int, default=0)
    parser.add_argument('--resume', action="store_true")
    parser.add_argument("--save-interval", type=int, default=4)
    args = parser.parse_known_args()[0]
    return args

# @pytest.mark.skipif(envpool is None, reason="EnvPool doesn't support this platform")
def ppo(args):
    # if you want to use python vector env, please refer to other test scripts
    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]
    if args.reward_threshold is None:
        default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250}
        args.reward_threshold = default_reward_threshold.get(
            args.task, env.spec.reward_threshold
        )
    # you can also use tianshou.env.SubprocVectorEnv
    # seed
    train_envs = DummyVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)]
    )
    # test_envs = gym.make(args.task)
    # test_envs = DummyVectorEnv(
    #     [lambda: gym.make(args.task) for _ in range(args.test_num)]
    # )
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    # model
    layer_num = 2
    actor = RecurrentActorProb(
        layer_num=layer_num,
        state_shape=args.state_shape,
        action_shape=args.action_shape,
        max_action=args.max_action,
        device=args.device,
        unbounded=True
    ).to(args.device)
    # actor_optim = torch.optim.Adam(actor.parameters(), lr=args.lr)
    hidden_layer_size = 64
    # critic = RecurrentCritic(
    #     layer_num=layer_num,
    #     state_shape=args.state_shape,
    #     action_shape=args.action_shape,
    #     hidden_layer_size=hidden_layer_size,
    #     device=args.device).to(args.device)
    critic = Critic(
        Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device),
        device=args.device
    ).to(args.device)
    actor_critic = ActorCritic(actor, critic)
    # actor_critic = torch.optim.Adam(critic.parameters(), lr=args.lr)

    # orthogonal initialization
    for m in actor_critic.modules():
        if isinstance(m, torch.nn.Linear):
            torch.nn.init.orthogonal_(m.weight)
            torch.nn.init.zeros_(m.bias)
    optim = torch.optim.Adam(actor_critic.parameters(), lr=args.lr)

    # replace DiagGuassian with Independent(Normal) which is equivalent
    # pass *logits to be consistent with policy.forward
    def dist(*logits):
        return Independent(Normal(*logits), 1)

    policy = PPOPolicy(
        actor,
        critic,
        optim,
        dist,
        discount_factor=args.gamma,
        max_grad_norm=args.max_grad_norm,
        eps_clip=args.eps_clip,
        vf_coef=args.vf_coef,
        ent_coef=args.ent_coef,
        reward_normalization=args.rew_norm,
        advantage_normalization=args.norm_adv,
        recompute_advantage=args.recompute_adv,
        dual_clip=args.dual_clip,
        value_clip=args.value_clip,
        gae_lambda=args.gae_lambda,
        action_space=env.action_space,
    )
    # collector
    train_collector = Collector(
        policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs))
    )
    # test_collector = Collector(policy, test_envs)
    test_collector=None
    # log
    log_path = os.path.join(args.logdir, args.task, "ppo")
    writer = SummaryWriter(log_path)
    logger = TensorboardLogger(writer, save_interval=args.save_interval)

    def save_best_fn(policy):
        torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))

    def stop_fn(mean_rewards):
        return mean_rewards >= args.reward_threshold

    def save_checkpoint_fn(epoch, env_step, gradient_step):
        # see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
        ckpt_path = os.path.join(log_path, "checkpoint.pth")
        # Example: saving by epoch num
        # ckpt_path = os.path.join(log_path, f"checkpoint_{epoch}.pth")
        torch.save(
            {
                "model": policy.state_dict(),
                "optim": optim.state_dict(),
            }, ckpt_path
        )
        return ckpt_path

    if args.resume:
        # load from existing checkpoint
        print(f"Loading agent under {log_path}")
        ckpt_path = os.path.join(log_path, "checkpoint.pth")
        if os.path.exists(ckpt_path):
            checkpoint = torch.load(ckpt_path, map_location=args.device)
            policy.load_state_dict(checkpoint["model"])
            optim.load_state_dict(checkpoint["optim"])
            print("Successfully restore policy and optim.")
        else:
            print("Fail to restore policy and optim.")

    # trainer
    trainer = OnpolicyTrainer(
        policy,
        train_collector,
        test_collector,
        args.epoch,
        args.step_per_epoch,
        args.repeat_per_collect,
        args.test_num,
        args.batch_size,
        episode_per_collect=args.episode_per_collect,
        stop_fn=stop_fn,
        save_best_fn=save_best_fn,
        logger=logger,
        resume_from_log=args.resume,
        save_checkpoint_fn=save_checkpoint_fn,
    )

    for epoch, epoch_stat, info in trainer:
        print(f"Epoch: {epoch}")
        print(epoch_stat)
        print(info)

    assert stop_fn(info["best_reward"])

    if __name__ == "__main__":
        pprint.pprint(info)
        # Let's watch its performance!
        env = gym.make(args.task)
        policy.eval()
        collector = Collector(policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        rews, lens = result["rews"], result["lens"]
        print(f"Final reward: {rews.mean()}, length: {lens.mean()}")

if __name__ == '__main__':
    args1 = get_args()
    ppo(args1)

[jaak-s]

Did you resolve the issue @1900360 ?

One thing that seems to be broken is that you are not using RecurrentCritic, as it is commented out:

    # critic = RecurrentCritic(
    #     layer_num=layer_num,
    #     state_shape=args.state_shape,
    #     action_shape=args.action_shape,
    #     hidden_layer_size=hidden_layer_size,
    #     device=args.device).to(args.device)