File size: 6,772 Bytes
23a59ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
from copy import deepcopy

from gymnasium.vector import AsyncVectorEnv, SyncVectorEnv
import gymnasium as gym
import numpy as np
import torch
from tensordict import TensorDict


MAX_OBS_DIM = 128
MAX_ACTION_DIM = 16


class VecWrapper(gym.Wrapper):
	"""
	A simple wrapper that unifies observation and action spaces across vectorized environments.
 	This is useful for environments that have different observation/action spaces per task.
	"""
	def __init__(self, env):
		super().__init__(env)
		self.orig_observation_space = env.observation_space
		self.orig_action_space = env.action_space
		if isinstance(env.observation_space, gym.spaces.Dict):  # State + RGB
			self.observation_space['state'] = gym.spaces.Box(
				low=-np.inf, high=np.inf, shape=(MAX_OBS_DIM,), dtype=np.float32)
		elif len(env.observation_space.shape) == 1:  # State
			self.observation_space = gym.spaces.Box(
				low=-np.inf, high=np.inf, shape=(MAX_OBS_DIM,), dtype=np.float32)
		else:  # RGB
			self.observation_space = env.observation_space
		self.action_space = gym.spaces.Box(
			low=-1, high=1, shape=(MAX_ACTION_DIM,), dtype=np.float32)

	def _pad_obs(self, obs):
		if isinstance(self.observation_space, gym.spaces.Dict):  # State + RGB
			if len(obs['state'].shape) == 1 and obs['state'].shape != self.observation_space['state'].shape:
				pad = np.zeros(self.observation_space['state'].shape[0] - obs['state'].shape[0], dtype=obs['state'].dtype)
				obs['state'] = np.concatenate((obs['state'], pad), axis=-1)
		elif len(self.observation_space.shape) == 1 and obs.shape != self.observation_space.shape:
			pad = np.zeros(self.observation_space.shape[0] - obs.shape[0], dtype=obs.dtype)
			obs = np.concatenate((obs, pad), axis=-1)
		return obs

	def reset(self, **kwargs):
		obs, info = self.env.reset(**kwargs)
		return self._pad_obs(obs), info

	def step(self, action):
		obs, reward, terminated, truncated, info = self.env.step(action[:self.orig_action_space.shape[0]])
		return self._pad_obs(obs), reward, terminated, truncated, info

	def render(self, *args, **kwargs):
		return self.env.render(*args, **kwargs).copy()

	def close(self):
		return self.env.close()


class VectorizedMultitaskWrapper(gym.Wrapper):
	"""
	Wrapper for vectorized multi-task environments.
	Each environment in the vectorized setup may have different observation and action spaces,
	and corresponds to different tasks. Creates one environment per task.
	"""

	def __init__(self, cfg, make_fn):
		self.cfg = deepcopy(cfg)
		self.cfg.task_embeddings = []  # Saves memory by not storing copies of embeddings
		
		# Create configs
		cfgs = [deepcopy(self.cfg) for _ in range(self.cfg.num_envs)]
		for i in range(len(cfgs)):
			cfgs[i].task = cfgs[i].tasks[i%self.cfg.num_tasks]
			cfgs[i].tasks = [cfgs[i].task]
			cfgs[i].num_envs = 1
			cfgs[i].child_env = True
			cfgs[i].seed = cfgs[i].seed + np.random.randint(1000)
		
		# Create environment
		wrapper = AsyncVectorEnv if self.cfg.get('env_mode', 'sync') == 'async' else SyncVectorEnv
		env_fns = [lambda c=cfgs[i]: VecWrapper(make_fn(c)) for i in range(self.cfg.num_envs)]
		self.env = wrapper(env_fns)
		super().__init__(self.env)
		
		# Define obs and action spaces
		if self.cfg.obs == 'state':
			self.observation_space = gym.spaces.Box(
				low=-np.inf, high=np.inf, shape=(MAX_OBS_DIM,), dtype=np.float32)
		else:
			self.observation_space = gym.spaces.Dict({
				'state': gym.spaces.Box(
					low=-np.inf, high=np.inf, shape=(MAX_OBS_DIM,), dtype=np.float32),
				'rgb': gym.spaces.Box(
					low=0, high=255, shape=(3, self.cfg.render_size, self.cfg.render_size), dtype=np.uint8),
			})
		self.action_space = gym.spaces.Box(
			low=-1, high=1, shape=(MAX_ACTION_DIM,), dtype=np.float32)
		assert self.action_space.shape[0] == self.env.action_space.shape[1], \
			"Action space mismatch between multitask wrapper and individual envs."		
		self._max_episode_steps = self.cfg.episode_lengths
		self.max_episode_steps = max(self._max_episode_steps)
		if self.cfg.rank == 0:
			print('Episode lengths:', self._max_episode_steps)
			print('Action dims:', self.cfg.action_dims)

	def rand_act(self):
		return torch.rand((self.cfg.num_envs, *self.action_space.shape)) * 2 - 1
	
	def _preprocess_obs(self, obs):
		if self.cfg.obs == 'state':
			obs = torch.tensor(obs, dtype=torch.float32)
		else:  # State + RGB
			obs = {
				'state': torch.tensor(obs['state'], dtype=torch.float32),
				'rgb': torch.tensor(obs['rgb'], dtype=torch.uint8),
			}
		return obs

	def reset(self):
		obs, info = self.env.reset()
		return self._preprocess_obs(obs), self._preprocess_info(info)
	
	def _preprocess_info(self, info):
		if 'final_info' in info:  # Handle final transitions
			assert 'final_observation' in info, \
				'Expected final observation in info when final_info is present.'
			fp64_to_fp32 = lambda x: x.astype(np.float32) if isinstance(x, np.ndarray) and x.dtype == np.float64 else x
			np_to_torch = lambda x: torch.from_numpy(fp64_to_fp32(x)) if isinstance(x, np.ndarray) else TensorDict({k: np_to_torch(v) for k, v in x.items()})
			info['final_observation'] = torch.stack([np_to_torch(d) for d in info['final_observation'] if d is not None])
			if self.cfg.save_video and self.cfg.get('num_demos', 0) > 0:
				info['final_frame'] = torch.stack([torch.from_numpy(d['frame']) for d in info['final_info'] if d is not None])
			keys = next(d.keys() for d in info['final_info'] if d is not None)
			pad = lambda vals: torch.from_numpy(
				np.stack([np.asarray(v, dtype=np.float32) if v is not None
					else np.full_like(next(x for x in vals if x is not None), np.nan, dtype=np.float32) for v in vals]))
			info['final_info'] = {
				k: pad([d[k] if d is not None else None for d in info['final_info']]) for k in keys}
		info['success'] = torch.tensor(info['success'], dtype=torch.float32)
		if 'frame' in info:
			info['frame'] = torch.stack([torch.from_numpy(d) for d in info['frame']])
		return info

	def step(self, action):
		obs, reward, terminated, truncated, info = self.env.step(action.numpy())
		return self._preprocess_obs(obs), \
			   torch.tensor(reward, dtype=torch.float32), \
			   torch.tensor(terminated, dtype=torch.bool), \
			   torch.tensor(truncated, dtype=torch.bool), \
			   self._preprocess_info(info)
	
	def render(self, *args, **kwargs):
		frames = []
		for env in self.env.envs:
			frame = env.render(*args, **kwargs)
			if frame is not None:
				frames.append(torch.from_numpy(frame))
		return torch.cat(frames, dim=1) if len(frames) > 0 else None


def make_vectorized_multitask_env(cfg, make_fn):
	"""Make a vectorized multi-task environment for world-model experiments."""
	print(f'[Rank {cfg.rank}] Creating multi-task environment with tasks: {cfg.tasks}')
	env = VectorizedMultitaskWrapper(cfg, make_fn)
	return env