import argparse import os import pickle import gymnasium as gym import numpy as np import torch from torch.utils.tensorboard import SummaryWriter from tianshou.data import Collector, PrioritizedVectorReplayBuffer, VectorReplayBuffer from tianshou.env import DummyVectorEnv from tianshou.policy import QRDQNPolicy from tianshou.policy.base import BasePolicy from tianshou.policy.modelfree.qrdqn import QRDQNTrainingStats from tianshou.trainer import OffpolicyTrainer from tianshou.utils import TensorboardLogger from tianshou.utils.net.common import Net from tianshou.utils.space_info import SpaceInfo def expert_file_name() -> str: return os.path.join(os.path.dirname(__file__), "expert_QRDQN_CartPole-v0.pkl") def get_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("--task", type=str, default="CartPole-v0") parser.add_argument("--reward-threshold", type=float, default=None) parser.add_argument("--seed", type=int, default=1) parser.add_argument("--eps-test", type=float, default=0.05) parser.add_argument("--eps-train", type=float, default=0.1) 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=0.9) parser.add_argument("--num-quantiles", type=int, default=200) parser.add_argument("--n-step", type=int, default=3) parser.add_argument("--target-update-freq", type=int, default=320) parser.add_argument("--epoch", type=int, default=10) parser.add_argument("--step-per-epoch", type=int, default=10000) parser.add_argument("--step-per-collect", type=int, default=10) parser.add_argument("--update-per-step", type=float, default=0.1) parser.add_argument("--batch-size", type=int, default=64) parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[128, 128, 128, 128]) parser.add_argument("--training-num", type=int, default=10) 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("--prioritized-replay", action="store_true", default=False) parser.add_argument("--alpha", type=float, default=0.6) parser.add_argument("--beta", type=float, default=0.4) parser.add_argument("--save-buffer-name", type=str, default=expert_file_name()) parser.add_argument( "--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu", ) return parser.parse_known_args()[0] def gather_data() -> VectorReplayBuffer | PrioritizedVectorReplayBuffer: args = get_args() env = gym.make(args.task) assert isinstance(env.action_space, gym.spaces.Discrete) 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-v0": 190} 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, action_shape=args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, softmax=False, num_atoms=args.num_quantiles, ) optim = torch.optim.Adam(net.parameters(), lr=args.lr) policy: QRDQNPolicy[QRDQNTrainingStats] = QRDQNPolicy( model=net, optim=optim, action_space=env.action_space, discount_factor=args.gamma, num_quantiles=args.num_quantiles, estimation_step=args.n_step, target_update_freq=args.target_update_freq, ).to(args.device) # buffer buf: VectorReplayBuffer | PrioritizedVectorReplayBuffer if args.prioritized_replay: buf = PrioritizedVectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), alpha=args.alpha, beta=args.beta, ) else: buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs)) # collector train_collector = Collector(policy, train_envs, buf, exploration_noise=True) train_collector.reset() test_collector = Collector(policy, test_envs, exploration_noise=True) test_collector.reset() # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num, reset_before_collect=True) # log log_path = os.path.join(args.logdir, args.task, "qrdqn") 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 def train_fn(epoch: int, env_step: int) -> None: # eps annnealing, just a demo if env_step <= 10000: policy.set_eps(args.eps_train) elif env_step <= 50000: eps = args.eps_train - (env_step - 10000) / 40000 * (0.9 * args.eps_train) policy.set_eps(eps) else: policy.set_eps(0.1 * args.eps_train) def test_fn(epoch: int, env_step: int | None) -> None: policy.set_eps(args.eps_test) # trainer result = OffpolicyTrainer( policy=policy, train_collector=train_collector, test_collector=test_collector, max_epoch=args.epoch, step_per_epoch=args.step_per_epoch, step_per_collect=args.step_per_collect, episode_per_test=args.test_num, batch_size=args.batch_size, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_best_fn=save_best_fn, logger=logger, update_per_step=args.update_per_step, ).run() assert stop_fn(result.best_reward) # save buffer in pickle format, for imitation learning unittest buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(test_envs)) policy.set_eps(0.2) collector = Collector(policy, test_envs, buf, exploration_noise=True) collector.reset() collector_stats = collector.collect(n_step=args.buffer_size) if args.save_buffer_name.endswith(".hdf5"): buf.save_hdf5(args.save_buffer_name) else: with open(args.save_buffer_name, "wb") as f: pickle.dump(buf, f) print(collector_stats) return buf