| import random |
| from argparse import Namespace |
| from typing import Callable |
|
|
| import gymnasium as gym |
| import numpy as np |
| import torch |
|
|
|
|
| def evaluate( |
| model_path: str, |
| make_env: Callable, |
| env_id: str, |
| eval_episodes: int, |
| run_name: str, |
| Model: torch.nn.Module, |
| device: torch.device = torch.device("cpu"), |
| epsilon: float = 0.05, |
| capture_video: bool = True, |
| ): |
| envs = gym.vector.SyncVectorEnv([make_env(env_id, 0, 0, capture_video, run_name)]) |
| model_data = torch.load(model_path, map_location="cpu") |
| args = Namespace(**model_data["args"]) |
| model = Model(envs, n_atoms=args.n_atoms, v_min=args.v_min, v_max=args.v_max) |
| model.load_state_dict(model_data["model_weights"]) |
| model = model.to(device) |
| model.eval() |
|
|
| obs, _ = envs.reset() |
| episodic_returns = [] |
| while len(episodic_returns) < eval_episodes: |
| if random.random() < epsilon: |
| actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)]) |
| else: |
| actions, _ = model.get_action(torch.Tensor(obs).to(device)) |
| actions = actions.cpu().numpy() |
| next_obs, _, _, _, infos = envs.step(actions) |
| if "final_info" in infos: |
| for info in infos["final_info"]: |
| if "episode" not in info: |
| continue |
| print(f"eval_episode={len(episodic_returns)}, episodic_return={info['episode']['r']}") |
| episodic_returns += [info["episode"]["r"]] |
| obs = next_obs |
|
|
| return episodic_returns |
|
|
|
|
| if __name__ == "__main__": |
| from huggingface_hub import hf_hub_download |
|
|
| from cleanrl.c51 import QNetwork, make_env |
|
|
| model_path = hf_hub_download(repo_id="cleanrl/CartPole-v1-c51-seed1", filename="c51.cleanrl_model") |
| evaluate( |
| model_path, |
| make_env, |
| "CartPole-v1", |
| eval_episodes=10, |
| run_name=f"eval", |
| Model=QNetwork, |
| device="cpu", |
| capture_video=False, |
| ) |
|
|
