forked from goncabakar/GMVAE-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
166 lines (130 loc) · 6.93 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
import torch
from torch import nn, optim
from torch.nn import functional as F
from fcnet import *
import math
class GMVAE(nn.Module):
def __init__(self, K, sigma, input_dim, x_dim, w_dim, hidden_dim, hidden_layers, device):
super(GMVAE, self).__init__()
self.K = K
self.sigma = sigma
self.input_dim = input_dim
self.x_dim = x_dim
self.w_dim = w_dim
self.hidden_dim = hidden_dim
self.hidden_layers = hidden_layers
self.device = device
# Q_xw
self.fc_q1 = FCNet(input_dim = self.input_dim, output_dim = self.hidden_dim, hidden_dim = self.hidden_dim, hidden_layers = self.hidden_layers, act_out = None)
self.fc_mean_x = nn.Linear(self.hidden_dim, self.x_dim)
self.fc_var_x = nn.Linear(self.hidden_dim, self.x_dim)
self.fc_mean_w = nn.Linear(self.hidden_dim, self.w_dim)
self.fc_var_w = nn.Linear(self.hidden_dim, self.w_dim)
self.fc_qz = nn.Linear(self.hidden_dim, self.K)
# Qz_x
self.softmax_qz = nn.Softmax(dim=1)
# Px_wz
self.fc_x_wz = FCNet(input_dim = self.w_dim, output_dim = self.hidden_dim, hidden_dim = self.hidden_dim, hidden_layers = 0, act_out = "tanh")
self.fc_x_means = nn.ModuleList()
self.fc_x_vars = nn.ModuleList()
self.x_mean_list = list()
self.x_var_list = list()
for i in range(self.K):
self.fc_x_means.append(nn.Linear(self.hidden_dim, self.x_dim))
self.fc_x_vars.append(nn.Linear(self.hidden_dim, self.x_dim))
# Py_x
self.fc_pyx = FCNet(input_dim = self.x_dim, output_dim = self.input_dim*2, hidden_dim = self.hidden_dim, hidden_layers = self.hidden_layers, act_out = None)
def Q_xw(self, y):
h = self.fc_q1(y)
mean_x = self.fc_mean_x(h)
var_x = torch.exp(self.fc_var_x(h))
mean_w = self.fc_mean_w(h)
var_w = torch.exp(self.fc_var_w(h))
#qz = self.softmax_qz(self.fc_qz(h))
#qz = F.softmax(self.fc_qz(h), dim=1)
return mean_x, var_x, mean_w, var_w
def Px_wz(self, w):
h2 = self.fc_x_wz(w)
self.x_mean_list = []
self.x_var_list = []
for i, l in enumerate(self.fc_x_means):
self.x_mean_list.append(l(h2))
for i, l in enumerate(self.fc_x_vars):
a = l(h2)
self.x_var_list.append(torch.exp(l(h2)))
return self.x_mean_list, self.x_var_list
def Py_x(self, x):
params = self.fc_pyx(x)
mean_y = params[:, 0:self.input_dim]
var_y = torch.exp(params[:, self.input_dim:])
return mean_y, var_y
def reparameterize(self, mu, var, dim1, dim2):
eps = torch.randn(dim1, dim2).to(self.device)
return mu + eps*torch.sqrt(var)
def loss_function(self, y, labels):
mean_x, var_x, mean_w, var_w = self.Q_xw(y)
w_sample = self.reparameterize(mu = mean_w, var = var_w, dim1 = mean_w.size()[0], dim2 = mean_w.size()[1])
x_sample = self.reparameterize(mu = mean_x, var = var_x, dim1 = mean_x.size()[0], dim2 = mean_x.size()[1])
y_recons_mean, y_recons_var = self.Py_x(x_sample)
x_mean_list, x_var_list = self.Px_wz(w_sample)
qz = self.Qz_x(x_mean_list, x_var_list)
predicted_labels = qz.max(1)[1]
predicted_labels = predicted_labels.float()
cov = (torch.sum( (predicted_labels - torch.mean(predicted_labels)) * (labels.view(-1) - torch.mean(labels.view(-1))))) * 1/(predicted_labels.size()[0]-1)
corr = cov/(torch.std(predicted_labels)*torch.std(labels))
# accuracy = torch.sum(labels.view(-1) == predicted_labels)
y_recons = self.reparameterize(mu = y_recons_mean, var = y_recons_var, dim1 = y_recons_mean.size()[0], dim2 = y_recons_mean.size()[1])
recons_loss = torch.sum(torch.pow(y - y_recons, 2),0)
recons_loss = recons_loss / y.size()[0]
#recons_loss = self.KL_recons_loss(x_sample, y_recons_mean, y_recons_var)
reg_w_loss = self.KL_gaussian_loss(mean_w, var_w)
reg_z_loss = self.KL_uniform_loss(qz)
reg_cond_loss = self.KL_conditional_loss(qz, mean_x, var_x, x_mean_list, x_var_list)
total_loss = recons_loss + reg_w_loss + reg_z_loss + reg_cond_loss
return recons_loss, reg_w_loss, reg_z_loss, reg_cond_loss, total_loss, corr
def KL_recons_loss(self, x_samples, y_recons_mean, y_recons_var):
loss = torch.zeros(1, requires_grad=True, device = self.device)
logvar = torch.log(y_recons_var)
loss = torch.sum(torch.sum(-logvar - torch.pow(x_samples - y_recons_mean, 2)/(2*torch.pow(y_recons_var, 2)), 1), 0)
#loss = 0.5 * torch.sum(torch.sum(var + torch.pow(mean, 2) - 1 - logvar, 1), 0)
return loss / (x_samples.size()[0]*x_samples.size()[1])
def KL_gaussian_loss(self, mean, var):
loss = torch.zeros(1, requires_grad=True, device = self.device)
logvar = torch.log(var)
loss = 0.5 * torch.sum(torch.sum(var + torch.pow(mean, 2) - 1 - logvar, 1), 0)
return loss / (mean.size()[0]*mean.size()[1])
def KL_uniform_loss(self, qz):
loss = torch.zeros(1, requires_grad=True, device = self.device)
for k in range(self.K):
loss = loss + torch.sum(qz[:,k] * (torch.log(self.K * qz[:,k] + 1e-10)),0)
return loss / (qz.size()[0]*qz.size()[1])
def KL_conditional_loss(self, qz, mean_x, var_x, x_mean_list, x_var_list):
# KL = 1/2( logvar2 - logvar1 + (var1 + (m1-m2)^2)/var2 - 1 )
x_mean_stack = torch.stack(x_mean_list)
x_var_stack = torch.stack(x_var_list)
K, bs, num_sample = x_mean_stack.size()
loss = torch.zeros(1, requires_grad=True, device = self.device)
for i in range(num_sample):
x_mean_2 = x_mean_stack[:,:,i].view(bs, K)
x_mean_1 = mean_x[:,i].view(bs, -1).repeat(1, K)
x_var_2 = x_var_stack[:,:,i].view(bs, K)
x_var_1 = var_x[:,i].view(bs, -1).repeat(1, K)
KL_batch = 0.5 * (torch.log(x_var_2) - torch.log(x_var_1) - 1 + (x_var_1 + torch.pow(x_mean_1 - x_mean_2, 2))/x_var_2)
weighted_KL = torch.sum(KL_batch*qz, 1)
loss = loss + torch.sum(weighted_KL,0)/weighted_KL.size()[0]
return loss / num_sample
def Qz_x(self, x_mean_list, x_var_list):
# KL = 1/2( logvar2 - logvar1 + (var1 + (m1-m2)^2)/var2 - 1 )
x_mean_stack = torch.stack(x_mean_list)
x_var_stack = torch.stack(x_var_list)
K, bs, num_sample = x_mean_stack.size()
qz = torch.zeros(bs, K, requires_grad=True, device = self.device)
for i in range(num_sample):
x_mean = x_mean_stack[:,:,i].view(bs, K)
x_var = x_var_stack[:,:,i].view(bs, K)
x_sample = self.reparameterize(mu = x_mean, var = x_var, dim1 = x_mean.size()[0], dim2 = x_mean.size()[1])
qz = qz + x_sample/K
#qz = qz/(torch.sum(x_sample, 1).view(bs,-1)*num_sample)
qz = self.softmax_qz(qz/num_sample)
return qz
return 0