-
Notifications
You must be signed in to change notification settings - Fork 5.5k
/
model_vector_env.py
164 lines (140 loc) · 5.76 KB
/
model_vector_env.py
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
import logging
from gymnasium.spaces import Discrete
import numpy as np
from ray.rllib.utils.annotations import override
from ray.rllib.env.vector_env import VectorEnv
from ray.rllib.evaluation.rollout_worker import get_global_worker
from ray.rllib.env.base_env import BaseEnv, convert_to_base_env
from ray.rllib.utils.typing import EnvType
logger = logging.getLogger(__name__)
def model_vector_env(env: EnvType) -> BaseEnv:
"""Returns a VectorizedEnv wrapper around the given environment.
To obtain worker configs, one can call get_global_worker().
Args:
env: The input environment (of any supported environment
type) to be convert to a _VectorizedModelGymEnv (wrapped as
an RLlib BaseEnv).
Returns:
BaseEnv: The BaseEnv converted input `env`.
"""
worker = get_global_worker()
worker_index = worker.worker_index
if worker_index:
env = _VectorizedModelGymEnv(
make_env=worker.make_sub_env_fn,
existing_envs=[env],
num_envs=worker.config.num_envs_per_worker,
observation_space=env.observation_space,
action_space=env.action_space,
)
return convert_to_base_env(
env,
make_env=worker.make_sub_env_fn,
num_envs=worker.config.num_envs_per_worker,
remote_envs=False,
remote_env_batch_wait_ms=0,
)
class _VectorizedModelGymEnv(VectorEnv):
"""Vectorized Environment Wrapper for MB-MPO.
Primary change is in the `vector_step` method, which calls the dynamics
models for next_obs "calculation" (instead of the actual env). Also, the
actual envs need to have two extra methods implemented: `reward(obs)` and
(optionally) `done(obs)`. If `done` is not implemented, we will assume
that episodes in the env do not terminate, ever.
"""
def __init__(
self,
make_env=None,
existing_envs=None,
num_envs=1,
*,
observation_space=None,
action_space=None,
env_config=None
):
self.make_env = make_env
self.envs = existing_envs
self.num_envs = num_envs
while len(self.envs) < num_envs:
self.envs.append(self.make_env(len(self.envs)))
self._timesteps = [0 for _ in range(self.num_envs)]
self.cur_obs = [None for _ in range(self.num_envs)]
super().__init__(
observation_space=observation_space or self.envs[0].observation_space,
action_space=action_space or self.envs[0].action_space,
num_envs=num_envs,
)
worker = get_global_worker()
self.model, self.device = worker.foreach_policy(
lambda x, y: (x.dynamics_model, x.device)
)[0]
@override(VectorEnv)
def vector_reset(self, *, seeds=None, options=None):
"""Override parent to store actual env obs for upcoming predictions."""
seeds = seeds or [None] * self.num_envs
options = options or [None] * self.num_envs
reset_results = [
e.reset(seed=seeds[i], options=options[i]) for i, e in enumerate(self.envs)
]
self.cur_obs = [io[0] for io in reset_results]
infos = [io[1] for io in reset_results]
self._timesteps = [0 for _ in range(self.num_envs)]
return self.cur_obs, infos
@override(VectorEnv)
def reset_at(self, index, *, seed=None, options=None):
"""Override parent to store actual env obs for upcoming predictions."""
obs, infos = self.envs[index].reset(seed=seed, options=options)
self.cur_obs[index] = obs
self._timesteps[index] = 0
return obs, infos
@override(VectorEnv)
def vector_step(self, actions):
if self.cur_obs is None:
raise ValueError("Need to reset env first")
for idx in range(self.num_envs):
self._timesteps[idx] += 1
# If discrete, need to one-hot actions
if isinstance(self.action_space, Discrete):
act = np.array(actions)
new_act = np.zeros((act.size, act.max() + 1))
new_act[np.arange(act.size), act] = 1
actions = new_act.astype("float32")
# Batch the TD-model prediction.
obs_batch = np.stack(self.cur_obs, axis=0)
action_batch = np.stack(actions, axis=0)
# Predict the next observation, given previous a) real obs
# (after a reset), b) predicted obs (any other time).
next_obs_batch = self.model.predict_model_batches(
obs_batch, action_batch, device=self.device
)
next_obs_batch = np.clip(next_obs_batch, -1000, 1000)
# Call env's reward function.
# Note: Each actual env must implement one to output exact rewards.
rew_batch = self.envs[0].reward(obs_batch, action_batch, next_obs_batch)
# If env has a `done` method, use it.
if hasattr(self.envs[0], "done"):
dones_batch = self.envs[0].done(next_obs_batch)
# Our sub-environments have timestep limits.
elif hasattr(self.envs[0], "_max_episode_steps"):
dones_batch = np.array(
[
self._timesteps[idx] >= self.envs[0]._max_episode_steps
for idx in range(self.num_envs)
]
)
# Otherwise, assume the episode does not end.
else:
dones_batch = np.asarray([False for _ in range(self.num_envs)])
truncateds_batch = [False for _ in range(self.num_envs)]
info_batch = [{} for _ in range(self.num_envs)]
self.cur_obs = next_obs_batch
return (
list(next_obs_batch),
list(rew_batch),
list(dones_batch),
truncateds_batch,
info_batch,
)
@override(VectorEnv)
def get_sub_environments(self):
return self.envs