reinforcement_learning: replace with unified-shape pools (6 cells x 100 checkpoints), full registry, configs, prompts
7d3c387 | """v2 统一形状 RL 池子的评测。不碰 RPG 里 v1 的评测链。 | |
| 用法: rl_eval_v2.py <alg> <env> <ckpt...> (ckpt 也可以是 'POOLMEAN:<池子目录>') | |
| """ | |
| import glob, os, sys | |
| import numpy as np, torch | |
| sys.path.insert(0, "/data/chl0045/project/benchmark_v2") | |
| import gymnasium as gym # 已被 sitecustomize 打过补丁(补观测维/Discrete(4)) | |
| from models.rl_actor_critic import create_actor_critic | |
| from models.rl_qnet import create_qnet | |
| alg, env_id = sys.argv[1], sys.argv[2] | |
| targets = sys.argv[3:] | |
| THR = {"Acrobot-v1": -100.0, "LunarLander-v3": 200.0}[env_id] | |
| def build(): | |
| if alg == "ppo": return create_actor_critic(obs_dim=8, n_actions=4, hidden_dim=64) | |
| return create_qnet(obs_dim=8, n_actions=4, hidden_dim=64) | |
| def load_sd(p): | |
| sd = torch.load(p, map_location="cpu") | |
| return sd.get("model_state_dict", sd) if isinstance(sd, dict) else sd | |
| def pool_mean(d): | |
| fs = sorted(glob.glob(d + "/**/*.pth", recursive=True)) | |
| acc = None | |
| for f in fs: | |
| sd = load_sd(f) | |
| if acc is None: acc = {k: v.double().clone() for k, v in sd.items()} | |
| else: | |
| for k in acc: acc[k] += sd[k].double() | |
| return {k: (v / len(fs)).float() for k, v in acc.items()}, len(fs) | |
| def run(sd, n_ep=int(__import__("os").environ.get("RL_EPISODES", 10)), seed0=12345): | |
| m = build(); m.load_state_dict(sd, strict=True); m.eval() | |
| head = m.actor if hasattr(m, "actor") else m.net | |
| e = gym.make(env_id) | |
| Rs, Ls = [], [] | |
| for i in range(n_ep): | |
| o, _ = e.reset(seed=seed0 + i); done = False; R = 0.0; L = 0 | |
| while not done: | |
| t = torch.from_numpy(np.asarray(o, dtype=np.float32)).unsqueeze(0) | |
| a = int(head(t).argmax(dim=-1).item()) | |
| o, r, te, tr, _ = e.step(a); R += r; L += 1; done = te or tr | |
| Rs.append(R); Ls.append(L) | |
| return float(np.mean(Rs)), float(np.std(Rs)), float(np.mean([r >= THR for r in Rs])), float(np.mean(Ls)) | |
| out = [] | |
| for t in targets: | |
| if t.startswith("POOLMEAN:"): | |
| sd, n = pool_mean(t.split(":", 1)[1]); name = f"池均值({n}个)" | |
| else: | |
| sd = load_sd(t); name = os.path.basename(t) | |
| R, S, succ, L = run(sd) | |
| out.append(R) | |
| print(f" {name:<26} R = {R:+9.2f} (std {S:6.2f}) succ={succ:.2f} len={L:.0f}", flush=True) | |
| if len(out) > 1: | |
| print(f" >>> 汇总 n={len(out)} R = {np.mean(out):+.2f} ± {np.std(out, ddof=1):.2f}", flush=True) | |