-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
209 lines (152 loc) · 6.56 KB
/
models.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
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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class CNNModel(nn.Module):
def __init__(self, in_channels=4, n_actions=14):
super(CNNModel, self).__init__()
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=8, stride=4)
self.conv2 = nn.Conv2d(32, 64, kernel_size=4, stride=2)
self.conv3 = nn.Conv2d(64, 64, kernel_size=3, stride=1)
self.fc4 = nn.Linear(3136, 512)
self.head = nn.Linear(512, n_actions)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = x.reshape(x.shape[0],-1)
x = F.relu(self.fc4(x))
return self.head(x)
class DuelingCNNModel(nn.Module):
def __init__(self, in_channels=4, n_actions=14):
super(DuelingCNNModel, self).__init__()
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=8, stride=4)
self.conv2 = nn.Conv2d(32, 64, kernel_size=4, stride=2)
self.conv3 = nn.Conv2d(64, 64, kernel_size=3, stride=1)
# Value net
self.value_fc = nn.Linear(3136, 512)
self.value_pred = nn.Linear(512, 1)
# Advantage net
self.advantage_fc = nn.Linear(3136, 512)
self.advantage_pred = nn.Linear(512, n_actions)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = x.reshape(x.shape[0],-1)
# Compute value
value_val = self.value_pred(F.relu(self.value_fc(x)))
# Compute advantage
advantage_val = self.advantage_pred(F.relu(self.advantage_fc(x)))
# Return estimated q val
return value_val + (advantage_val - advantage_val.mean())
class NoisyNet(nn.Module):
"""
Re-formatted but originally from: https://github.com/higgsfield/RL-Adventure/blob/master/5.noisy%20dqn.ipynb
"""
def __init__(self, in_dim, out_dim, std=0.5):
super(NoisyNet, self).__init__()
self.in_dim = in_dim
self.out_dim = out_dim
self.std = std
# Weight for linear layer
self.mean_weight = nn.Parameter(torch.FloatTensor(out_dim, in_dim))
self.std_weight = nn.Parameter(torch.FloatTensor(out_dim, in_dim))
# Bias for linear layer
self.mean_bias = nn.Parameter(torch.FloatTensor(out_dim))
self.std_bias = nn.Parameter(torch.FloatTensor(out_dim))
# Register as buffer
self.register_buffer('weight_epsilon', torch.FloatTensor(out_dim, in_dim))
self.register_buffer('bias_epsilon', torch.FloatTensor(out_dim))
# Reset values
self.reset_params()
self.reset_noise()
def forward(self, x):
if self.training:
w = self.mean_weight + self.std_weight.mul(self.weight_epsilon)
b = self.mean_bias + self.std_bias.mul(self.bias_epsilon)
else:
w = self.mean_weight
b = self.mean_bias
return F.linear(x, w, b)
def reset_params(self):
# Set mean range
mean_range = 1/math.sqrt(self.mean_weight.size(1))
# Fill weight and bias data
self.mean_weight.data.uniform_(-mean_range, mean_range)
self.std_weight.data.fill_(self.std / math.sqrt(self.std_weight.size(1)))
self.mean_bias.data.uniform_(-mean_range, mean_range)
self.std_bias.data.fill_(self.std / math.sqrt(self.std_bias.size(0)))
def reset_noise(self):
epsilon_in = self.scale_noise(self.in_dim)
epsilon_out = self.scale_noise(self.out_dim)
self.weight_epsilon.copy_(epsilon_out.ger(epsilon_in))
self.bias_epsilon.copy_(self.scale_noise(self.out_dim))
def scale_noise(self, size):
x = torch.randn(size)
x = x.sign().mul(x.abs().sqrt())
return x
class NoisyDuelingCNNModel(nn.Module):
def __init__(self, in_channels=4, n_actions=14):
super(NoisyDuelingCNNModel, self).__init__()
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=8, stride=4)
self.conv2 = nn.Conv2d(32, 64, kernel_size=4, stride=2)
self.conv3 = nn.Conv2d(64, 64, kernel_size=3, stride=1)
# Value noisy net
self.value_fc = NoisyNet(3136, 512)
self.value_pred = NoisyNet(512, 1)
# Advantage noisy net
self.advantage_fc = NoisyNet(3136, 512)
self.advantage_pred = NoisyNet(512, n_actions)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = x.reshape(x.shape[0],-1)
# Compute value
value_val = self.value_pred(F.relu(self.value_fc(x)))
# Compute advantage
advantage_val = self.advantage_pred(F.relu(self.advantage_fc(x)))
# Return estimated q val
return value_val + (advantage_val - advantage_val.mean())
def reset_noise(self):
self.value_fc.reset_noise()
self.value_pred.reset_noise()
self.advantage_fc.reset_noise()
self.advantage_pred.reset_noise()
class DistributionalNoisyDuelingCNNModel(nn.Module):
def __init__(self, number_atoms,in_channels=4, n_actions=14):
super(DistributionalNoisyDuelingCNNModel, self).__init__()
self.number_atoms = number_atoms
self.n_actions = n_actions
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=8, stride=4)
self.conv2 = nn.Conv2d(32, 64, kernel_size=4, stride=2)
self.conv3 = nn.Conv2d(64, 64, kernel_size=3, stride=1)
# Value noisy net
self.value_fc = NoisyNet(3136, 512)
self.value_pred = NoisyNet(512, self.number_atoms)
# Advantage noisy net
self.advantage_fc = NoisyNet(3136, 512)
self.advantage_pred = NoisyNet(512, n_actions*self.number_atoms)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = x.reshape(x.shape[0],-1)
# Compute advantage
advantage_val = self.advantage_pred(F.relu(self.advantage_fc(x)))
advantage_val = advantage_val.view(-1, self.n_actions, self.number_atoms)
# Compute value
value_val = self.value_pred(F.relu(self.value_fc(x)))
value_val = value_val.view(-1, 1, self.number_atoms)
# Compute q distribution
q_atoms = value_val + (advantage_val - advantage_val.mean(dim=1, keepdim=True))
# Take softmax
q_atoms = F.softmax(q_atoms,dim=2)
return q_atoms
def reset_noise(self):
self.value_fc.reset_noise()
self.value_pred.reset_noise()
self.advantage_fc.reset_noise()
self.advantage_pred.reset_noise()
#