File size: 5,585 Bytes
8796ba9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

from __future__ import annotations

from dataclasses import dataclass
from typing import Optional, Tuple

import torch
import torch.nn as nn
from torch.distributions import Normal

from transformers import PreTrainedModel
from transformers.utils import ModelOutput

from .configuration_mrbalance import MrBalanceConfig


@dataclass
class MrBalanceOutput(ModelOutput):
    """

    Outputs returned by MrBalance.

    """

    action: Optional[torch.FloatTensor] = None
    action_mean: Optional[torch.FloatTensor] = None
    action_std: Optional[torch.FloatTensor] = None
    value: Optional[torch.FloatTensor] = None
    log_prob: Optional[torch.FloatTensor] = None


class MrBalanceMLPForRL(PreTrainedModel):
    """

    Hugging Face-compatible MrBalance Actor-Critic network.



    Exact architecture:



        64 -> 128 -> 128 -> 64



    with:



        actor:  64 -> 2

        critic: 64 -> 1



    and:



        learned log_std: 2

    """

    config_class = MrBalanceConfig
    base_model_prefix = "mrbalance"
    main_input_name = "observation"

    def __init__(self, config: MrBalanceConfig):
        super().__init__(config)

        self.backbone = nn.Sequential(
            nn.Linear(
                config.observation_size,
                config.hidden_size,
            ),
            nn.SiLU(),

            nn.Linear(
                config.hidden_size,
                config.intermediate_size,
            ),
            nn.SiLU(),

            nn.Linear(
                config.intermediate_size,
                config.bottleneck_size,
            ),
            nn.SiLU(),
        )

        self.actor = nn.Linear(
            config.bottleneck_size,
            config.action_size,
        )

        self.critic = nn.Linear(
            config.bottleneck_size,
            1,
        )

        self.log_std = nn.Parameter(
            torch.full(
                (config.action_size,),
                config.actor_log_std_init,
            )
        )

        self.post_init()

    def _init_weights(self, module: nn.Module) -> None:
        """

        Match the original MrBalance initialization.



        These values are only relevant for a newly initialized model.

        During export, the learned checkpoint weights overwrite them.

        """

        if not isinstance(module, nn.Linear):
            return

        if module is self.actor:
            gain = 0.01
        elif module is self.critic:
            gain = 1.0
        else:
            gain = 2.0 ** 0.5

        nn.init.orthogonal_(
            module.weight,
            gain=gain,
        )

        if module.bias is not None:
            nn.init.zeros_(module.bias)

    def _stats(

        self,

        observation: torch.Tensor,

    ) -> Tuple[
        torch.Tensor,
        torch.Tensor,
        torch.Tensor,
    ]:
        hidden = self.backbone(observation)

        mean = self.actor(hidden)

        value = self.critic(hidden).squeeze(-1)

        std = self.log_std.exp().expand_as(mean)

        return mean, std, value

    def get_value(

        self,

        observation: torch.Tensor,

    ) -> torch.Tensor:
        _, _, value = self._stats(observation)
        return value

    def get_action_and_value(

        self,

        observation: torch.Tensor,

        raw_action: Optional[torch.Tensor] = None,

        deterministic: bool = False,

    ):
        mean, std, value = self._stats(observation)

        dist = Normal(mean, std)

        if raw_action is None:
            raw_action = (
                mean
                if deterministic
                else dist.sample()
            )

        action = torch.tanh(raw_action)

        # Exact same Tanh log-probability correction
        # as the original PPO implementation.
        log_prob = dist.log_prob(
            raw_action
        ).sum(-1)

        log_prob -= torch.log(
            torch.clamp(
                1.0 - action.pow(2),
                min=1e-6,
            )
        ).sum(-1)

        entropy = dist.entropy().sum(-1)

        return (
            action,
            log_prob,
            entropy,
            value,
            raw_action,
        )

    def forward(

        self,

        observation: torch.Tensor,

        deterministic: bool = False,

        raw_action: Optional[torch.Tensor] = None,

        return_dict: bool = True,

        **kwargs,

    ):
        action, log_prob, _, value, _ = (
            self.get_action_and_value(
                observation=observation,
                raw_action=raw_action,
                deterministic=deterministic,
            )
        )

        mean, std, _ = self._stats(observation)

        if not return_dict:
            return (
                action,
                mean,
                std,
                value,
                log_prob,
            )

        return MrBalanceOutput(
            action=action,
            action_mean=mean,
            action_std=std,
            value=value,
            log_prob=log_prob,
        )

    @torch.no_grad()
    def predict_action(

        self,

        observation: torch.Tensor,

        deterministic: bool = True,

    ) -> torch.Tensor:
        output = self.forward(
            observation,
            deterministic=deterministic,
        )
        return output.action


MrBalanceModel = MrBalanceMLPForRL