forked from iShohei220/torch-gqn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
230 lines (183 loc) · 7.67 KB
/
model.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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
import torch
from torch import nn
from torch.nn import functional as F
from torch.distributions import Normal
from torch.distributions.kl import kl_divergence
from representation import Pyramid, Tower, Pool
from core import InferenceCore, GenerationCore
class GQN(nn.Module):
def __init__(self, representation="pool", L=12, shared_core=False):
super(GQN, self).__init__()
# Number of generative layers
self.L = L
# Representation network
self.representation = representation
if representation=="pyramid":
self.phi = Pyramid()
elif representation=="tower":
self.phi = Tower()
elif representation=="pool":
self.phi = Pool()
# Generation network
self.shared_core = shared_core
if shared_core:
self.inference_core = InferenceCore()
self.generation_core = GenerationCore()
else:
self.inference_core = nn.ModuleList([InferenceCore() for _ in range(L)])
self.generation_core = nn.ModuleList([GenerationCore() for _ in range(L)])
self.eta_pi = nn.Conv2d(128, 2*3, kernel_size=5, stride=1, padding=2)
self.eta_g = nn.Conv2d(128, 3, kernel_size=1, stride=1, padding=0)
self.eta_e = nn.Conv2d(128, 2*3, kernel_size=5, stride=1, padding=2)
# EstimateELBO
def forward(self, x, v, v_q, x_q, sigma):
B, M, *_ = x.size()
# Scene encoder
if self.representation=='tower':
r = x.new_zeros((B, 256, 16, 16))
else:
r = x.new_zeros((B, 256, 1, 1))
for k in range(M):
r_k = self.phi(x[:, k], v[:, k])
r += r_k
# Generator initial state
c_g = x.new_zeros((B, 128, 16, 16))
h_g = x.new_zeros((B, 128, 16, 16))
u = x.new_zeros((B, 128, 64, 64))
# Inference initial state
c_e = x.new_zeros((B, 128, 16, 16))
h_e = x.new_zeros((B, 128, 16, 16))
elbo = 0
for l in range(self.L):
# Prior factor
mu_pi, logvar_pi = torch.split(self.eta_pi(h_g), 3, dim=1)
std_pi = torch.exp(0.5*logvar_pi)
pi = Normal(mu_pi, std_pi)
# Inference state update
if self.shared_core:
c_e, h_e = self.inference_core(x_q, v_q, r, c_e, h_e, h_g, u)
else:
c_e, h_e = self.inference_core[l](x_q, v_q, r, c_e, h_e, h_g, u)
# Posterior factor
mu_q, logvar_q = torch.split(self.eta_e(h_e), 3, dim=1)
std_q = torch.exp(0.5*logvar_q)
q = Normal(mu_q, std_q)
# Posterior sample
z = q.rsample()
# Generator state update
if self.shared_core:
c_g, h_g, u = self.generation_core(v_q, r, c_g, h_g, u, z)
else:
c_g, h_g, u = self.generation_core[l](v_q, r, c_g, h_g, u, z)
# ELBO KL contribution update
elbo -= torch.sum(kl_divergence(q, pi), dim=[1,2,3])
# ELBO likelihood contribution update
elbo += torch.sum(Normal(self.eta_g(u), sigma).log_prob(x_q), dim=[1,2,3])
return elbo
def generate(self, x, v, v_q):
B, M, *_ = x.size()
# Scene encoder
if self.representation=='tower':
r = x.new_zeros((B, 256, 16, 16))
else:
r = x.new_zeros((B, 256, 1, 1))
for k in range(M):
r_k = self.phi(x[:, k], v[:, k])
r += r_k
# Initial state
c_g = x.new_zeros((B, 128, 16, 16))
h_g = x.new_zeros((B, 128, 16, 16))
u = x.new_zeros((B, 128, 64, 64))
for l in range(self.L):
# Prior factor
mu_pi, logvar_pi = torch.split(self.eta_pi(h_g), 3, dim=1)
std_pi = torch.exp(0.5*logvar_pi)
pi = Normal(mu_pi, std_pi)
# Prior sample
z = pi.sample()
# State update
if self.shared_core:
c_g, h_g, u = self.generation_core(v_q, r, c_g, h_g, u, z)
else:
c_g, h_g, u = self.generation_core[l](v_q, r, c_g, h_g, u, z)
# Image sample
mu = self.eta_g(u)
return torch.clamp(mu, 0, 1)
def kl_divergence(self, x, v, v_q, x_q):
B, M, *_ = x.size()
# Scene encoder
if self.representation=='tower':
r = x.new_zeros((B, 256, 16, 16))
else:
r = x.new_zeros((B, 256, 1, 1))
for k in range(M):
r_k = self.phi(x[:, k], v[:, k])
r += r_k
# Generator initial state
c_g = x.new_zeros((B, 128, 16, 16))
h_g = x.new_zeros((B, 128, 16, 16))
u = x.new_zeros((B, 128, 64, 64))
# Inference initial state
c_e = x.new_zeros((B, 128, 16, 16))
h_e = x.new_zeros((B, 128, 16, 16))
kl = 0
for l in range(self.L):
# Prior factor
mu_pi, logvar_pi = torch.split(self.eta_pi(h_g), 3, dim=1)
std_pi = torch.exp(0.5*logvar_pi)
pi = Normal(mu_pi, std_pi)
# Inference state update
if self.shared_core:
c_e, h_e = self.inference_core(x_q, v_q, r, c_e, h_e, h_g, u)
else:
c_e, h_e = self.inference_core[l](x_q, v_q, r, c_e, h_e, h_g, u)
# Posterior factor
mu_q, logvar_q = torch.split(self.eta_e(h_e), 3, dim=1)
std_q = torch.exp(0.5*logvar_q)
q = Normal(mu_q, std_q)
# Posterior sample
z = q.rsample()
# Generator state update
if self.shared_core:
c_g, h_g, u = self.generation_core(v_q, r, c_g, h_g, u, z)
else:
c_g, h_g, u = self.generation_core[l](v_q, r, c_g, h_g, u, z)
# ELBO KL contribution update
kl += torch.sum(kl_divergence(q, pi), dim=[1,2,3])
return kl
def reconstruct(self, x, v, v_q, x_q):
B, M, *_ = x.size()
# Scene encoder
if self.representation=='tower':
r = x.new_zeros((B, 256, 16, 16))
else:
r = x.new_zeros((B, 256, 1, 1))
for k in range(M):
r_k = self.phi(x[:, k], v[:, k])
r += r_k
# Generator initial state
c_g = x.new_zeros((B, 128, 16, 16))
h_g = x.new_zeros((B, 128, 16, 16))
u = x.new_zeros((B, 128, 64, 64))
# Inference initial state
c_e = x.new_zeros((B, 128, 16, 16))
h_e = x.new_zeros((B, 128, 16, 16))
for l in range(self.L):
# Inference state update
if self.shared_core:
c_e, h_e = self.inference_core(x_q, v_q, r, c_e, h_e, h_g, u)
else:
c_e, h_e = self.inference_core[l](x_q, v_q, r, c_e, h_e, h_g, u)
# Posterior factor
mu_q, logvar_q = torch.split(self.eta_e(h_e), 3, dim=1)
std_q = torch.exp(0.5*logvar_q)
q = Normal(mu_q, std_q)
# Posterior sample
z = q.rsample()
# Generator state update
if self.shared_core:
c_g, h_g, u = self.generation_core(v_q, r, c_g, h_g, u, z)
else:
c_g, h_g, u = self.generation_core[l](v_q, r, c_g, h_g, u, z)
mu = self.eta_g(u)
return torch.clamp(mu, 0, 1)