import argparse import os import gymnasium as gym import numpy as np import torch from torch.distributions import Distribution, Independent, Normal 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, Net from tianshou.utils.net.continuous import ActorProb, Critic from tianshou.utils.space_info import SpaceInfo def get_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("--task", type=str, default="Pendulum-v1") 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=1e-3) parser.add_argument("--gamma", type=float, default=0.95) parser.add_argument("--epoch", type=int, default=5) parser.add_argument("--step-per-epoch", type=int, default=150000) parser.add_argument("--episode-per-collect", type=int, default=16) parser.add_argument("--repeat-per-collect", type=int, default=2) parser.add_argument("--batch-size", type=int, default=128) parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[64, 64]) parser.add_argument("--training-num", type=int, default=16) 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=0.25) 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) 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 args.max_action = space_info.action_info.max_action 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 if env.spec else None, ) # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) 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 = ActorProb(net, args.action_shape, unbounded=True, device=args.device).to(args.device) critic = Critic( Net(state_shape=args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device), 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) # replace DiagGuassian with Independent(Normal) which is equivalent # pass *logits to be consistent with policy.forward def dist(loc_scale: tuple[torch.Tensor, torch.Tensor]) -> Distribution: loc, scale = loc_scale return Independent(Normal(loc, scale), 1) policy: PPOPolicy[PPOTrainingStats] = PPOPolicy( actor=actor, critic=critic, optim=optim, dist_fn=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) # 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: 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 def save_checkpoint_fn(epoch: int, env_step: int, gradient_step: int) -> str: # 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=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, 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_stat in trainer: print(f"Epoch: {epoch_stat.epoch}") print(epoch_stat) # print(info) assert stop_fn(epoch_stat.info_stat.best_reward) def test_ppo_resume(args: argparse.Namespace = get_args()) -> None: args.resume = True test_ppo(args)