import argparse import os import gymnasium as gym import numpy as np import torch import torch.nn as nn from gymnasium.spaces import Box from torch.utils.tensorboard import SummaryWriter from tianshou.data import Collector, VectorReplayBuffer from tianshou.env import DummyVectorEnv from tianshou.policy import PPOPolicy from tianshou.policy.base import BasePolicy from tianshou.policy.modelfree.ppo import PPOTrainingStats from tianshou.trainer import OnpolicyTrainer from tianshou.utils import TensorboardLogger from tianshou.utils.net.common import ActorCritic, DataParallelNet, Net from tianshou.utils.net.discrete import Actor, Critic from tianshou.utils.space_info import SpaceInfo def get_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("--task", type=str, default="CartPole-v1") parser.add_argument("--reward-threshold", type=float, default=None) parser.add_argument("--seed", type=int, default=1626) parser.add_argument("--buffer-size", type=int, default=20000) parser.add_argument("--lr", type=float, default=3e-4) parser.add_argument("--gamma", type=float, default=0.99) parser.add_argument("--epoch", type=int, default=10) parser.add_argument("--step-per-epoch", type=int, default=50000) parser.add_argument("--step-per-collect", type=int, default=2000) parser.add_argument("--repeat-per-collect", type=int, default=10) parser.add_argument("--batch-size", type=int, default=64) parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[64, 64]) 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=0.0) 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=0.5) 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=0) parser.add_argument("--norm-adv", type=int, default=0) parser.add_argument("--recompute-adv", type=int, default=0) parser.add_argument("--dual-clip", type=float, default=None) parser.add_argument("--value-clip", type=int, default=0) return parser.parse_known_args()[0] def test_ppo(args: argparse.Namespace = get_args()) -> None: env = gym.make(args.task) space_info = SpaceInfo.from_env(env) args.state_shape = space_info.observation_info.obs_shape args.action_shape = space_info.action_info.action_shape if args.reward_threshold is None: default_reward_threshold = {"CartPole-v1": 195} args.reward_threshold = default_reward_threshold.get( args.task, env.spec.reward_threshold if env.spec else None, ) # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(state_shape=args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor: nn.Module critic: nn.Module if torch.cuda.is_available(): actor = DataParallelNet(Actor(net, args.action_shape, device=args.device).to(args.device)) critic = DataParallelNet(Critic(net, device=args.device).to(args.device)) else: actor = Actor(net, args.action_shape, device=args.device).to(args.device) critic = Critic(net, device=args.device).to(args.device) actor_critic = ActorCritic(actor, critic) # 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) dist = torch.distributions.Categorical policy: PPOPolicy[PPOTrainingStats] = PPOPolicy( actor=actor, critic=critic, optim=optim, dist_fn=dist, action_scaling=isinstance(env.action_space, Box), 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, gae_lambda=args.gae_lambda, reward_normalization=args.rew_norm, dual_clip=args.dual_clip, value_clip=args.value_clip, action_space=env.action_space, deterministic_eval=True, advantage_normalization=args.norm_adv, recompute_advantage=args.recompute_adv, ) # collector train_collector = Collector( policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), ) test_collector = Collector(policy, test_envs) # log log_path = os.path.join(args.logdir, args.task, "ppo") writer = SummaryWriter(log_path) logger = TensorboardLogger(writer) def save_best_fn(policy: BasePolicy) -> None: torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth")) def stop_fn(mean_rewards: float) -> bool: return mean_rewards >= args.reward_threshold # trainer result = OnpolicyTrainer( policy=policy, train_collector=train_collector, test_collector=test_collector, max_epoch=args.epoch, step_per_epoch=args.step_per_epoch, repeat_per_collect=args.repeat_per_collect, episode_per_test=args.test_num, batch_size=args.batch_size, step_per_collect=args.step_per_collect, stop_fn=stop_fn, save_best_fn=save_best_fn, logger=logger, ).run() assert stop_fn(result.best_reward)