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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362 | from typing import Any, Optional, Union
from gymnasium import spaces
import torch
from torch import nn
from lit_sb3.core.preprocessing import get_action_dim
from lit_sb3.core.types import PyTorchObs, Schedule
from lit_sb3.modules import (
BaseFeaturesExtractor,
CombinedExtractor,
FlattenExtractor,
NatureCNN,
create_mlp,
get_actor_critic_arch,
)
from lit_sb3.policies import BasePolicy, ContinuousCritic
class Actor(BasePolicy):
"""Actor network (policy) for TD3.
Args:
observation_space: Observation space
action_space: Action space
net_arch: Network architecture
features_extractor: Network to extract features
(a CNN when using images, a nn.Flatten() layer otherwise)
features_dim: Number of features
activation_fn: Activation function
normalize_images: Whether to normalize images or not,
dividing by 255.0 (True by default)
"""
def __init__(
self,
observation_space: spaces.Space,
action_space: spaces.Box,
net_arch: list[int],
features_extractor: nn.Module,
features_dim: int,
activation_fn: type[nn.Module] = nn.ReLU,
normalize_images: bool = True,
) -> None:
super().__init__(
observation_space,
action_space,
features_extractor=features_extractor,
normalize_images=normalize_images,
squash_output=True,
)
self.net_arch = net_arch
self.features_dim = features_dim
self.activation_fn = activation_fn
action_dim = get_action_dim(self.action_space)
actor_net = create_mlp(features_dim, action_dim, net_arch, activation_fn, squash_output=True)
# Deterministic action
self.mu = nn.Sequential(*actor_net)
def _get_constructor_parameters(self) -> dict[str, Any]:
data = super()._get_constructor_parameters()
data.update(
{
"net_arch": self.net_arch,
"features_dim": self.features_dim,
"activation_fn": self.activation_fn,
"features_extractor": self.features_extractor,
}
)
return data
def forward(self, obs: torch.Tensor) -> torch.Tensor:
# assert deterministic, 'The TD3 actor only outputs deterministic actions'
features = self.extract_features(obs, self.features_extractor)
return self.mu(features)
def _predict(self, observation: PyTorchObs, deterministic: bool = False) -> torch.Tensor:
# Note: the deterministic deterministic parameter is ignored in the case of TD3.
# Predictions are always deterministic.
return self(observation)
class TD3Policy(BasePolicy):
"""Policy class (with both actor and critic) for TD3.
Args:
observation_space: Observation space
action_space: Action space
lr_schedule: Learning rate schedule (could be constant)
net_arch: The specification of the policy and value networks.
activation_fn: Activation function
features_extractor_class: Features extractor to use.
features_extractor_kwargs: Keyword arguments
to pass to the features extractor.
normalize_images: Whether to normalize images or not,
dividing by 255.0 (True by default)
optimizer_class: The optimizer to use,
``torch.optim.Adam`` by default
optimizer_kwargs: Additional keyword arguments,
excluding the learning rate, to pass to the optimizer
n_critics: Number of critic networks to create.
share_features_extractor: Whether to share or not the features extractor
between the actor and the critic (this saves computation time)
"""
actor: Actor
actor_target: Actor
critic: ContinuousCritic
critic_target: ContinuousCritic
def __init__(
self,
observation_space: spaces.Space,
action_space: spaces.Box,
lr_schedule: Schedule,
net_arch: list[int] | dict[str, list[int]] | None = None,
activation_fn: type[nn.Module] = nn.ReLU,
features_extractor_class: type[BaseFeaturesExtractor] = FlattenExtractor,
features_extractor_kwargs: dict[str, Any] | None = None,
normalize_images: bool = True,
optimizer_class: type[torch.optim.Optimizer] = torch.optim.Adam,
optimizer_kwargs: dict[str, Any] | None = None,
n_critics: int = 2,
share_features_extractor: bool = False,
) -> None:
super().__init__(
observation_space,
action_space,
features_extractor_class,
features_extractor_kwargs,
optimizer_class=optimizer_class,
optimizer_kwargs=optimizer_kwargs,
squash_output=True,
normalize_images=normalize_images,
)
# Default network architecture, from the original paper
if net_arch is None:
net_arch = [256, 256] if features_extractor_class == NatureCNN else [400, 300]
actor_arch, critic_arch = get_actor_critic_arch(net_arch)
self.net_arch = net_arch
self.activation_fn = activation_fn
self.net_args = {
"observation_space": self.observation_space,
"action_space": self.action_space,
"net_arch": actor_arch,
"activation_fn": self.activation_fn,
"normalize_images": normalize_images,
}
self.actor_kwargs = self.net_args.copy()
self.critic_kwargs = self.net_args.copy()
self.critic_kwargs.update(
{
"n_critics": n_critics,
"net_arch": critic_arch,
"share_features_extractor": share_features_extractor,
}
)
self.share_features_extractor = share_features_extractor
self._build(lr_schedule)
def _build(self, lr_schedule: Schedule) -> None:
# Create actor and target
# the features extractor should not be shared
self.actor = self.make_actor(features_extractor=None)
self.actor_target = self.make_actor(features_extractor=None)
# Initialize the target to have the same weights as the actor
self.actor_target.load_state_dict(self.actor.state_dict())
self.actor.optimizer = self.optimizer_class(
self.actor.parameters(),
lr=lr_schedule(1), # type: ignore[call-arg]
**self.optimizer_kwargs,
)
if self.share_features_extractor:
self.critic = self.make_critic(features_extractor=self.actor.features_extractor)
# Critic target should not share the features extractor with critic
# but it can share it with the actor target as actor and critic are sharing
# the same features_extractor too
# NOTE: as a result the effective poliak (soft-copy) coefficient for the features extractor
# will be 2 * tau instead of tau (updated one time with the actor, a second time with the critic)
self.critic_target = self.make_critic(features_extractor=self.actor_target.features_extractor)
else:
# Create new features extractor for each network
self.critic = self.make_critic(features_extractor=None)
self.critic_target = self.make_critic(features_extractor=None)
self.critic_target.load_state_dict(self.critic.state_dict())
self.critic.optimizer = self.optimizer_class(
self.critic.parameters(),
lr=lr_schedule(1), # type: ignore[call-arg]
**self.optimizer_kwargs,
)
# Target networks should always be in eval mode
self.actor_target.set_training_mode(False)
self.critic_target.set_training_mode(False)
def _get_constructor_parameters(self) -> dict[str, Any]:
data = super()._get_constructor_parameters()
data.update(
{
"net_arch": self.net_arch,
"activation_fn": self.net_args["activation_fn"],
"n_critics": self.critic_kwargs["n_critics"],
"lr_schedule": self._dummy_schedule, # dummy lr schedule, not needed for loading policy alone
"optimizer_class": self.optimizer_class,
"optimizer_kwargs": self.optimizer_kwargs,
"features_extractor_class": self.features_extractor_class,
"features_extractor_kwargs": self.features_extractor_kwargs,
"share_features_extractor": self.share_features_extractor,
}
)
return data
def make_actor(self, features_extractor: BaseFeaturesExtractor | None = None) -> Actor:
actor_kwargs = self._update_features_extractor(self.actor_kwargs, features_extractor)
return Actor(**actor_kwargs).to(self.device)
def make_critic(self, features_extractor: BaseFeaturesExtractor | None = None) -> ContinuousCritic:
critic_kwargs = self._update_features_extractor(self.critic_kwargs, features_extractor)
return ContinuousCritic(**critic_kwargs).to(self.device)
def forward(self, observation: PyTorchObs, deterministic: bool = False) -> torch.Tensor:
return self._predict(observation, deterministic=deterministic)
def _predict(self, observation: PyTorchObs, deterministic: bool = False) -> torch.Tensor:
# Note: the deterministic deterministic parameter is ignored in the case of TD3.
# Predictions are always deterministic.
return self.actor(observation)
def set_training_mode(self, mode: bool) -> None:
"""Put the policy in either training or evaluation mode.
This affects certain modules, such as batch normalisation and dropout.
Args:
mode: if true, set to training mode, else set to evaluation mode
"""
self.actor.set_training_mode(mode)
self.critic.set_training_mode(mode)
self.training = mode
MlpPolicy = TD3Policy
class CnnPolicy(TD3Policy):
"""Policy class (with both actor and critic) for TD3.
Args:
observation_space: Observation space
action_space: Action space
lr_schedule: Learning rate schedule (could be constant)
net_arch: The specification of the policy and value networks.
activation_fn: Activation function
features_extractor_class: Features extractor to use.
features_extractor_kwargs: Keyword arguments
to pass to the features extractor.
normalize_images: Whether to normalize images or not,
dividing by 255.0 (True by default)
optimizer_class: The optimizer to use,
``torch.optim.Adam`` by default
optimizer_kwargs: Additional keyword arguments,
excluding the learning rate, to pass to the optimizer
n_critics: Number of critic networks to create.
share_features_extractor: Whether to share or not the features extractor
between the actor and the critic (this saves computation time)
"""
def __init__(
self,
observation_space: spaces.Space,
action_space: spaces.Box,
lr_schedule: Schedule,
net_arch: list[int] | dict[str, list[int]] | None = None,
activation_fn: type[nn.Module] = nn.ReLU,
features_extractor_class: type[BaseFeaturesExtractor] = NatureCNN,
features_extractor_kwargs: dict[str, Any] | None = None,
normalize_images: bool = True,
optimizer_class: type[torch.optim.Optimizer] = torch.optim.Adam,
optimizer_kwargs: dict[str, Any] | None = None,
n_critics: int = 2,
share_features_extractor: bool = False,
) -> None:
super().__init__(
observation_space,
action_space,
lr_schedule,
net_arch,
activation_fn,
features_extractor_class,
features_extractor_kwargs,
normalize_images,
optimizer_class,
optimizer_kwargs,
n_critics,
share_features_extractor,
)
class MultiInputPolicy(TD3Policy):
"""Policy class (with both actor and critic) for TD3 to be used with Dict observation spaces.
Args:
observation_space: Observation space
action_space: Action space
lr_schedule: Learning rate schedule (could be constant)
net_arch: The specification of the policy and value networks.
activation_fn: Activation function
features_extractor_class: Features extractor to use.
features_extractor_kwargs: Keyword arguments
to pass to the features extractor.
normalize_images: Whether to normalize images or not,
dividing by 255.0 (True by default)
optimizer_class: The optimizer to use,
``torch.optim.Adam`` by default
optimizer_kwargs: Additional keyword arguments,
excluding the learning rate, to pass to the optimizer
n_critics: Number of critic networks to create.
share_features_extractor: Whether to share or not the features extractor
between the actor and the critic (this saves computation time)
"""
def __init__(
self,
observation_space: spaces.Dict,
action_space: spaces.Box,
lr_schedule: Schedule,
net_arch: list[int] | dict[str, list[int]] | None = None,
activation_fn: type[nn.Module] = nn.ReLU,
features_extractor_class: type[BaseFeaturesExtractor] = CombinedExtractor,
features_extractor_kwargs: dict[str, Any] | None = None,
normalize_images: bool = True,
optimizer_class: type[torch.optim.Optimizer] = torch.optim.Adam,
optimizer_kwargs: dict[str, Any] | None = None,
n_critics: int = 2,
share_features_extractor: bool = False,
) -> None:
super().__init__(
observation_space,
action_space,
lr_schedule,
net_arch,
activation_fn,
features_extractor_class,
features_extractor_kwargs,
normalize_images,
optimizer_class,
optimizer_kwargs,
n_critics,
share_features_extractor,
)
|