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 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter
from torch.distributions import Distribution
from torch.distributions import Categorical
"""
A Discrete NADE, as a torch Distribution object
"""
class DiscreteNADEDistribution(Distribution):
def __init__(self, data_size, data_domain_sizes, hidden_size, w, v, c):
self.data_size = data_size
self.data_domain_sizes = data_domain_sizes
self.hidden_size = hidden_size
self.w = w
self.v = v
self.c = c
self.temperature = 1
def set_temperature(self, temp):
self.temperature = temp
def log_prob(self, x):
log_probs = []
batch_size = x.size()[0]
# Make this have same batch size as x
a = self.c
a = a.unsqueeze(0)
a = a.expand(batch_size, a.size()[1])
for i in range(0, self.data_size):
h = F.sigmoid(a)
probs = F.softmax(self.v[i](h), 1)
dist = Categorical(probs)
val = x[:, i]
lp = dist.log_prob(val)
log_probs.append(lp.unsqueeze(1))
normalized_val = val.float() / self.data_domain_sizes[i] # Normalize to [0,1]
normalized_val = normalized_val.unsqueeze(1)
if i < self.data_size - 1:
a = self.w[i](normalized_val) + a
log_prob = torch.sum(torch.cat(log_probs, 1), 1) # First dimension is batch dim
return log_prob
def sample(self):
return self.sample_n(1)
def sample_n(self, n):
outputs = []
batch_size = n
# Make this have batch size
a = self.c
a = a.unsqueeze(0)
a = a.expand(batch_size, a.size()[1])
for i in range(0, self.data_size):
h = F.sigmoid(a)
logits = self.v[i](h)
probs = F.softmax(logits / self.temperature, 1)
dist = Categorical(probs)
val = dist.sample().unsqueeze(1)
outputs.append(val)
normalized_val = val.float() / self.data_domain_sizes[i] # Normalize to [0,1]
if i < self.data_size - 1:
a = self.w[i](normalized_val) + a
outputs = torch.cat(outputs, 1) # First dimension is batch dim
return outputs
"""
A Discrete NADE, as a nn Module
"""
class DiscreteNADEModule(nn.Module):
def __init__(self, data_size, data_domain_sizes, hidden_size):
super(DiscreteNADEModule, self).__init__()
self.data_size = data_size
self.data_domain_sizes = data_domain_sizes
self.hidden_size = hidden_size
# The initial bias
self.c = Parameter(torch.Tensor(hidden_size))
self.reset_parameters()
# Initialize one linear module for every step of the first layer
# (Need ModuleList to make automatic parameter registration work)
self.w = nn.ModuleList()
for i in range(0, data_size - 1):
self.w.append(nn.Linear(1, hidden_size, bias=False))
# Initialize one linear module for every step of the second layer
self.v = nn.ModuleList()
for i in range(0, data_size):
domain_size = data_domain_sizes[i]
self.v.append(nn.Linear(hidden_size, domain_size, bias=True))
self.temperature = 1
def reset_parameters(self):
stdv = 1. / math.sqrt(self.hidden_size)
self.c.data.uniform_(-stdv, stdv)
def set_temperature(self, temp):
self.temperature = temp
def forward(self, x):
return self.log_prob(x)
def log_prob(self, x):
nade = DiscreteNADEDistribution(self.data_size, self.data_domain_sizes, self.hidden_size,
self.w, self.v, self.c)
return nade.log_prob(x)
def sample(self):
nade = DiscreteNADEDistribution(self.data_size, self.data_domain_sizes, self.hidden_size,
self.w, self.v, self.c)
nade.set_temperature(self.temperature)
return nade.sample()
def sample_n(self, n):
nade = DiscreteNADEDistribution(self.data_size, self.data_domain_sizes, self.hidden_size,
self.w, self.v, self.c)
nade.set_temperature(self.temperature)
return nade.sample_n(n)
# ##### TEST
# n = DiscreteNADEModule(10, [3, 3, 3, 3, 3, 3, 3, 3, 3, 3], 5)
# # inp = torch.autograd.Variable(torch.LongTensor(4, 10).fill_(1))
# # lp = n(inp)
# # print(lp)
# out = n.sample_n(4)
# print(out)
# ##### END TEST
|