File size: 8,564 Bytes
f238a34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Any, Dict, List, Optional, Type

import gym
import torch as th
from torch import nn

from stable_baselines3.common.policies import BasePolicy, register_policy
from stable_baselines3.common.torch_layers import BaseFeaturesExtractor, FlattenExtractor, NatureCNN, create_mlp
from stable_baselines3.common.type_aliases import Schedule


class QNetwork(BasePolicy):
    """
    Action-Value (Q-Value) network for DQN

    :param observation_space: Observation space
    :param action_space: Action space
    :param net_arch: The specification of the policy and value networks.
    :param activation_fn: Activation function
    :param normalize_images: Whether to normalize images or not,
         dividing by 255.0 (True by default)
    """

    def __init__(
        self,
        observation_space: gym.spaces.Space,
        action_space: gym.spaces.Space,
        features_extractor: nn.Module,
        features_dim: int,
        net_arch: Optional[List[int]] = None,
        activation_fn: Type[nn.Module] = nn.ReLU,
        normalize_images: bool = True,
    ):
        super(QNetwork, self).__init__(
            observation_space,
            action_space,
            features_extractor=features_extractor,
            normalize_images=normalize_images,
        )

        if net_arch is None:
            net_arch = [64, 64]

        self.net_arch = net_arch
        self.activation_fn = activation_fn
        self.features_extractor = features_extractor
        self.features_dim = features_dim
        self.normalize_images = normalize_images
        action_dim = self.action_space.n  # number of actions
        q_net = create_mlp(self.features_dim, action_dim, self.net_arch, self.activation_fn)
        self.q_net = nn.Sequential(*q_net)

    def forward(self, obs: th.Tensor) -> th.Tensor:
        """
        Predict the q-values.

        :param obs: Observation
        :return: The estimated Q-Value for each action.
        """
        return self.q_net(self.extract_features(obs))

    def _predict(self, observation: th.Tensor, deterministic: bool = True) -> th.Tensor:
        q_values = self.forward(observation)
        # Greedy action
        action = q_values.argmax(dim=1).reshape(-1)
        return action

    def _get_constructor_parameters(self) -> Dict[str, Any]:
        data = super()._get_constructor_parameters()

        data.update(
            dict(
                net_arch=self.net_arch,
                features_dim=self.features_dim,
                activation_fn=self.activation_fn,
                features_extractor=self.features_extractor,
            )
        )
        return data


class DQNPolicy(BasePolicy):
    """
    Policy class with Q-Value Net and target net for DQN

    :param observation_space: Observation space
    :param action_space: Action space
    :param lr_schedule: Learning rate schedule (could be constant)
    :param net_arch: The specification of the policy and value networks.
    :param activation_fn: Activation function
    :param features_extractor_class: Features extractor to use.
    :param features_extractor_kwargs: Keyword arguments
        to pass to the features extractor.
    :param normalize_images: Whether to normalize images or not,
         dividing by 255.0 (True by default)
    :param optimizer_class: The optimizer to use,
        ``th.optim.Adam`` by default
    :param optimizer_kwargs: Additional keyword arguments,
        excluding the learning rate, to pass to the optimizer
    """

    def __init__(
        self,
        observation_space: gym.spaces.Space,
        action_space: gym.spaces.Space,
        lr_schedule: Schedule,
        net_arch: Optional[List[int]] = None,
        activation_fn: Type[nn.Module] = nn.ReLU,
        features_extractor_class: Type[BaseFeaturesExtractor] = FlattenExtractor,
        features_extractor_kwargs: Optional[Dict[str, Any]] = None,
        normalize_images: bool = True,
        optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
        optimizer_kwargs: Optional[Dict[str, Any]] = None,
    ):
        super(DQNPolicy, self).__init__(
            observation_space,
            action_space,
            features_extractor_class,
            features_extractor_kwargs,
            optimizer_class=optimizer_class,
            optimizer_kwargs=optimizer_kwargs,
        )

        if net_arch is None:
            if features_extractor_class == FlattenExtractor:
                net_arch = [64, 64]
            else:
                net_arch = []

        self.net_arch = net_arch
        self.activation_fn = activation_fn
        self.normalize_images = normalize_images

        self.net_args = {
            "observation_space": self.observation_space,
            "action_space": self.action_space,
            "net_arch": self.net_arch,
            "activation_fn": self.activation_fn,
            "normalize_images": normalize_images,
        }

        self.q_net, self.q_net_target = None, None
        self._build(lr_schedule)

    def _build(self, lr_schedule: Schedule) -> None:
        """
        Create the network and the optimizer.

        :param lr_schedule: Learning rate schedule
            lr_schedule(1) is the initial learning rate
        """

        self.q_net = self.make_q_net()
        self.q_net_target = self.make_q_net()
        self.q_net_target.load_state_dict(self.q_net.state_dict())

        # Setup optimizer with initial learning rate
        self.optimizer = self.optimizer_class(self.parameters(), lr=lr_schedule(1), **self.optimizer_kwargs)

    def make_q_net(self) -> QNetwork:
        # Make sure we always have separate networks for features extractors etc
        net_args = self._update_features_extractor(self.net_args, features_extractor=None)
        return QNetwork(**net_args).to(self.device)

    def forward(self, obs: th.Tensor, deterministic: bool = True) -> th.Tensor:
        return self._predict(obs, deterministic=deterministic)

    def _predict(self, obs: th.Tensor, deterministic: bool = True) -> th.Tensor:
        return self.q_net._predict(obs, deterministic=deterministic)

    def _get_constructor_parameters(self) -> Dict[str, Any]:
        data = super()._get_constructor_parameters()

        data.update(
            dict(
                net_arch=self.net_args["net_arch"],
                activation_fn=self.net_args["activation_fn"],
                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,
            )
        )
        return data


MlpPolicy = DQNPolicy


class CnnPolicy(DQNPolicy):
    """
    Policy class for DQN when using images as input.

    :param observation_space: Observation space
    :param action_space: Action space
    :param lr_schedule: Learning rate schedule (could be constant)
    :param net_arch: The specification of the policy and value networks.
    :param activation_fn: Activation function
    :param features_extractor_class: Features extractor to use.
    :param normalize_images: Whether to normalize images or not,
         dividing by 255.0 (True by default)
    :param optimizer_class: The optimizer to use,
        ``th.optim.Adam`` by default
    :param optimizer_kwargs: Additional keyword arguments,
        excluding the learning rate, to pass to the optimizer
    """

    def __init__(
        self,
        observation_space: gym.spaces.Space,
        action_space: gym.spaces.Space,
        lr_schedule: Schedule,
        net_arch: Optional[List[int]] = None,
        activation_fn: Type[nn.Module] = nn.ReLU,
        features_extractor_class: Type[BaseFeaturesExtractor] = NatureCNN,
        features_extractor_kwargs: Optional[Dict[str, Any]] = None,
        normalize_images: bool = True,
        optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
        optimizer_kwargs: Optional[Dict[str, Any]] = None,
    ):
        super(CnnPolicy, self).__init__(
            observation_space,
            action_space,
            lr_schedule,
            net_arch,
            activation_fn,
            features_extractor_class,
            features_extractor_kwargs,
            normalize_images,
            optimizer_class,
            optimizer_kwargs,
        )


register_policy("MlpPolicy", MlpPolicy)
register_policy("CnnPolicy", CnnPolicy)