reinforcement_learning: replace with unified-shape pools (6 cells x 100 checkpoints), full registry, configs, prompts
7d3c387 | """把观测维补齐到 RL_OBS_DIM、动作空间声明成 RL_N_ACT。 | |
| 不设这两个环境变量时什么都不做, 所以对现有脚本零影响。 | |
| 目的: 让 Acrobot(obs6/act3) 和 LunarLander(obs8/act4) 训出逐字节同形的网络。""" | |
| import os | |
| _OBS = os.environ.get("RL_OBS_DIM"); _ACT = os.environ.get("RL_N_ACT") | |
| if _OBS and _ACT: | |
| _OBS, _ACT = int(_OBS), int(_ACT) | |
| import numpy as np, gymnasium as gym | |
| class _Unify(gym.Wrapper): | |
| def __init__(self, env): | |
| super().__init__(env) | |
| self.n_valid = int(env.action_space.n) | |
| sp = env.observation_space | |
| self._pad = _OBS - int(sp.shape[0]) | |
| assert self._pad >= 0, f"RL_OBS_DIM={_OBS} 小于真实观测维 {sp.shape[0]}" | |
| z = np.zeros(self._pad, dtype=np.float32) | |
| self.observation_space = gym.spaces.Box( | |
| low=np.concatenate([sp.low.astype(np.float32), z - np.inf if False else z]), | |
| high=np.concatenate([sp.high.astype(np.float32), z]), | |
| dtype=np.float32) | |
| self.action_space = gym.spaces.Discrete(_ACT) | |
| def _o(self, o): | |
| return np.concatenate([np.asarray(o, dtype=np.float32), | |
| np.zeros(self._pad, dtype=np.float32)]) | |
| def reset(self, **kw): | |
| o, i = self.env.reset(**kw); return self._o(o), i | |
| def step(self, a): | |
| a = int(a) | |
| if a >= self.n_valid: # 掩码生效时不该发生; 留个安全网并计数 | |
| a = self.n_valid - 1 | |
| _Unify.n_clipped += 1 | |
| o, r, te, tr, i = self.env.step(a); return self._o(o), r, te, tr, i | |
| _Unify.n_clipped = 0 | |
| _orig = gym.make | |
| def _make(env_id, *a, **kw): | |
| return _Unify(_orig(env_id, *a, **kw)) | |
| gym.make = _make | |