-
Notifications
You must be signed in to change notification settings - Fork 0
/
modules.py
178 lines (142 loc) · 5.99 KB
/
modules.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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
from math import sqrt
import torch
from torch import nn
from torch.distributions import Distribution, Normal
import numpy as np
from typing import Tuple
class Actor(nn.Module):
def __init__(self,
state_dim: int,
action_dim: int,
hidden_dim: int = 256,
min_log_std: float = -20.0,
max_log_std: float = 2.0,
min_action: float = -1.0,
max_action: float = 1.0) -> None:
super().__init__()
self.net = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, action_dim)
)
self.log_std = nn.Parameter(torch.zeros(action_dim, dtype=torch.float32))
self.min_log_std = min_log_std
self.max_log_std = max_log_std
self.min_action = min_action
self.max_action = max_action
def get_policy(self, state: torch.Tensor) -> Distribution:
mean = self.net(state)
log_std = self.log_std.clamp(self.min_log_std, self.max_log_std)
policy = Normal(mean, log_std.exp())
return policy
def log_prob(self, state: torch.Tensor, action: torch.Tensor) -> torch.Tensor:
policy = self.get_policy(state)
log_prob = policy.log_prob(action).sum(-1, keepdim=True)
return log_prob
def act(self, state: np.ndarray, device: str) -> np.ndarray:
state = torch.tensor(state[None], dtype=torch.float32, device=device)
policy = self.get_policy(state)
#action = policy.mean
if self.net.training:
action = policy.sample()
else:
action = policy.mean
return action[0].cpu().numpy()
def forward(self, state: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
policy = self.get_policy(state)
action = policy.rsample().clamp(self.min_action, self.max_action)
log_prob = policy.log_prob(action).sum(-1, keepdim=True)
return action, log_prob
class EnsembledLinear(nn.Module):
def __init__(self,
in_features: int,
out_features: int,
ensemble_size: int) -> None:
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.ensemble_size = ensemble_size
self.weight = nn.Parameter(torch.empty(ensemble_size, in_features, out_features))
self.bias = nn.Parameter(torch.empty(ensemble_size, 1, out_features))
self.reset_parameters()
def reset_parameters(self) -> None:
scale_factor = sqrt(5)
# default pytorch init
for layer in range(self.ensemble_size):
nn.init.kaiming_normal_(self.weight[layer], a=scale_factor)
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight[0])
bound = 1 / sqrt(fan_in) if fan_in > 0 else 0
nn.init.uniform_(self.bias, -bound, bound)
def forward(self, x: torch.Tensor) -> torch.Tensor:
'''
x: [ensemble_size, batch_size, input_size]
weight: [ensemble_size, input_size, out_size]
bias: [ensemble_size, batch_size, out_size]
'''
return x @ self.weight + self.bias
class ValueFunction(nn.Module):
def __init__(self,
state_dim: int,
hidden_dim: int = 256,
layer_norm: bool = True) -> None:
super().__init__()
self.state_dim = state_dim
self.hidden_dim = hidden_dim
self.net = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
nn.Linear(hidden_dim, 1)
)
def forward(self, state: torch.Tensor) -> torch.Tensor:
return self.net(state)
class EnsembledCritic(nn.Module):
def __init__(self,
state_dim: int,
action_dim: int,
hidden_dim: int = 256,
num_critics: int = 2,
layer_norm: bool = True,
edac_init: bool = True) -> None:
super().__init__()
#block = nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity()
self.num_critics = num_critics
self.critic = nn.Sequential(
EnsembledLinear(state_dim + action_dim, hidden_dim, num_critics),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
EnsembledLinear(hidden_dim, hidden_dim, num_critics),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
EnsembledLinear(hidden_dim, hidden_dim, num_critics),
nn.LayerNorm(hidden_dim) if layer_norm else nn.Identity(),
nn.ReLU(),
EnsembledLinear(hidden_dim, 1, num_critics)
)
if edac_init:
# init as in the EDAC paper
for layer in self.critic[::3]:
nn.init.constant_(layer.bias, 0.1)
nn.init.uniform_(self.critic[-1].weight, -3e-3, 3e-3)
nn.init.uniform_(self.critic[-1].bias, -3e-3, 3e-3)
def forward(self, state: torch.Tensor, action: torch.Tensor) -> torch.Tensor:
concat = torch.cat([state, action], dim=-1)
concat = concat.unsqueeze(0)
concat = concat.repeat_interleave(self.num_critics, dim=0)
q_values = self.critic(concat).squeeze(-1)
return q_values
if __name__ == "__main__":
action = torch.rand(12, 6)
state = torch.rand(12, 17)
actor = Actor(17, 6)
print(actor.log_prob(state, action).shape)