| import numpy as np | |
| class Driver: | |
| def __init__(self, envs, **kwargs): | |
| self._envs = envs | |
| self._kwargs = kwargs | |
| self._on_steps = [] | |
| self._on_resets = [] | |
| self._on_episodes = [] | |
| self._act_spaces = [env.act_space for env in envs] | |
| self.reset() | |
| def on_step(self, callback): | |
| self._on_steps.append(callback) | |
| def on_reset(self, callback): | |
| self._on_resets.append(callback) | |
| def on_episode(self, callback): | |
| self._on_episodes.append(callback) | |
| def reset(self): | |
| self._obs = [None] * len(self._envs) | |
| self._eps = [None] * len(self._envs) | |
| self._state = None | |
| def __call__(self, policy, steps=0, episodes=0): | |
| step, episode = 0, 0 | |
| while step < steps or episode < episodes: | |
| obs = { | |
| i: self._envs[i].reset() | |
| for i, ob in enumerate(self._obs) if ob is None or ob['is_last']} | |
| for i, ob in obs.items(): | |
| self._obs[i] = ob() if callable(ob) else ob | |
| act = {k: np.zeros(v.shape) for k, v in self._act_spaces[i].items()} | |
| tran = {k: self._convert(v) for k, v in {**ob, **act}.items()} | |
| [fn(tran, worker=i, **self._kwargs) for fn in self._on_resets] | |
| self._eps[i] = [tran] | |
| obs = {k: np.stack([o[k] for o in self._obs]) for k in self._obs[0]} | |
| actions, self._state = policy(obs, self._state, **self._kwargs) | |
| actions = [ | |
| {k: np.array(actions[k][i]) for k in actions} | |
| for i in range(len(self._envs))] | |
| assert len(actions) == len(self._envs) | |
| obs = [e.step(a) for e, a in zip(self._envs, actions)] | |
| obs = [ob() if callable(ob) else ob for ob in obs] | |
| for i, (act, ob) in enumerate(zip(actions, obs)): | |
| tran = {k: self._convert(v) for k, v in {**ob, **act}.items()} | |
| [fn(tran, worker=i, **self._kwargs) for fn in self._on_steps] | |
| self._eps[i].append(tran) | |
| step += 1 | |
| if ob['is_last']: | |
| ep = self._eps[i] | |
| ep = {k: self._convert([t[k] for t in ep]) for k in ep[0]} | |
| [fn(ep, **self._kwargs) for fn in self._on_episodes] | |
| episode += 1 | |
| self._obs = obs | |
| def _convert(self, value): | |
| value = np.array(value) | |
| if np.issubdtype(value.dtype, np.floating): | |
| return value.astype(np.float32) | |
| elif np.issubdtype(value.dtype, np.signedinteger): | |
| return value.astype(np.int32) | |
| elif np.issubdtype(value.dtype, np.uint8): | |
| return value.astype(np.uint8) | |
| return value | |
| class MultiEnvDriver: | |
| def __init__(self, envs, modes, **kwargs): | |
| self._envs = envs | |
| self._kwargs = kwargs | |
| self._on_steps = [] | |
| self._on_resets = [] | |
| self._on_episodes = [] | |
| self._act_spaces = [env.act_space for env in envs] | |
| self.reset() | |
| self.modes = modes | |
| def on_step(self, callback): | |
| self._on_steps.append(callback) | |
| def on_reset(self, callback): | |
| self._on_resets.append(callback) | |
| def on_episode(self, callback): | |
| self._on_episodes.append(callback) | |
| def reset(self): | |
| self._obs = [None] * len(self._envs) | |
| self._eps = [None] * len(self._envs) | |
| self._state = None | |
| def __call__(self, policy, steps=0, episodes=0): | |
| step, episode = 0, 0 | |
| while step < steps or episode < episodes: | |
| obs = { | |
| i: self._envs[i].reset() | |
| for i, ob in enumerate(self._obs) if ob is None or ob['is_last']} | |
| for i, ob in obs.items(): | |
| self._obs[i] = ob() if callable(ob) else ob | |
| act = {k: np.zeros(v.shape) for k, v in self._act_spaces[i].items()} | |
| tran = {k: self._convert(v) for k, v in {**ob, **act}.items()} | |
| [fn(tran, worker=i, **self._kwargs) for fn in self._on_resets] | |
| self._eps[i] = [tran] | |
| obs = {k: np.stack([o[k] for o in self._obs]) for k in self._obs[0]} | |
| actions, self._state = policy(obs, self._state, **self._kwargs) | |
| actions = [ | |
| {k: np.array(actions[k][i]) for k in actions} | |
| for i in range(len(self._envs))] | |
| assert len(actions) == len(self._envs) | |
| obs = [e.step(a) for e, a in zip(self._envs, actions)] | |
| obs = [ob() if callable(ob) else ob for ob in obs] | |
| for i, (act, ob) in enumerate(zip(actions, obs)): | |
| tran = {k: self._convert(v) for k, v in {**ob, **act}.items()} | |
| [fn(tran, worker=i, **self._kwargs) for fn in self._on_steps] | |
| self._eps[i].append(tran) | |
| step += 1 | |
| if ob['is_last']: | |
| ep = self._eps[i] | |
| ep = {k: self._convert([t[k] for t in ep]) for k in ep[0]} | |
| [fn(ep, self.modes[i], **self._kwargs) for fn in self._on_episodes] | |
| episode += 1 | |
| self._obs = obs | |
| def _convert(self, value): | |
| value = np.array(value) | |
| if np.issubdtype(value.dtype, np.floating): | |
| return value.astype(np.float32) | |
| elif np.issubdtype(value.dtype, np.signedinteger): | |
| return value.astype(np.int32) | |
| elif np.issubdtype(value.dtype, np.uint8): | |
| return value.astype(np.uint8) | |
| return value | |