/
model.py
177 lines (142 loc) · 8.46 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
import torch.nn as nn
from torch.nn.parameter import Parameter
import torch
from utils import *
import numpy as np
import os
# from ../PGBN_tool import PGBN_sampler
import torch.nn.functional as F
import scipy.io as sio # mat
class GBN_model(nn.Module):
def __init__(self, args):
super(GBN_model, self).__init__()
self.real_min = torch.tensor(1e-30)
self.wei_shape_max = torch.tensor(10.0).float()
self.wei_shape = torch.tensor(1e-1).float()
self.vocab_size = args.vocab_size
self.hidden_size = args.hidden_size
self.topic_size = args.topic_size
self.topic_size = [self.vocab_size] + self.topic_size
self.layer_num = len(self.topic_size) - 1
self.embed_size = args.embed_size
self.bn_layer = nn.ModuleList([nn.BatchNorm1d(self.hidden_size) for i in range(self.layer_num)])
h_encoder = [Conv1D(self.hidden_size, 1, self.vocab_size)]
for i in range(self.layer_num - 1):
h_encoder.append(Conv1D(self.hidden_size, 1, self.hidden_size))
self.h_encoder = nn.ModuleList(h_encoder)
shape_encoder = [Conv1D(self.topic_size[i + 1], 1, self.topic_size[i + 1] + self.hidden_size) for i in
range(self.layer_num - 1)]
shape_encoder.append(Conv1D(self.topic_size[self.layer_num], 1, self.hidden_size))
self.shape_encoder = nn.ModuleList(shape_encoder)
scale_encoder = [Conv1D(self.topic_size[i + 1], 1, self.topic_size[i + 1] + self.hidden_size) for i in range(self.layer_num - 1)]
scale_encoder.append(Conv1D(self.topic_size[self.layer_num], 1, self.hidden_size))
self.scale_encoder = nn.ModuleList(scale_encoder)
decoder = [GaussSoftmaxV3(self.topic_size[i], self.topic_size[i + 1], self.embed_size) for i in
range(self.layer_num)]
self.decoder = nn.ModuleList(decoder)
for t in range(self.layer_num - 1):
self.decoder[t + 1].mu = self.decoder[t].mu_c
self.decoder[t + 1].log_sigma = self.decoder[t].log_sigma_c
graph_wordnet = sio.loadmat('./dataset/TopicTree_20ng.mat')
self.graph = []
for i in range(len(graph_wordnet['graph_topic_adj'][0])):
self.graph.append(torch.from_numpy(graph_wordnet['graph_topic_adj'][0][i]).cuda())
self.ob = 1.0
def log_max(self, x):
return torch.log(torch.max(x, self.real_min.cuda()))
def reparameterize(self, Wei_shape_res, Wei_scale, Sample_num=1):
# sample one
eps = torch.cuda.FloatTensor(Sample_num, Wei_shape_res.shape[0], Wei_shape_res.shape[1]).uniform_(0, 1)
theta = torch.unsqueeze(Wei_scale, axis=0).repeat(Sample_num, 1, 1) \
* torch.pow(-log_max(1 - eps), torch.unsqueeze(Wei_shape_res, axis=0).repeat(Sample_num, 1, 1)) #
theta = torch.max(theta, self.real_min.cuda())
return torch.mean(theta, dim=0, keepdim=False)
def reparameterize2(self, Wei_shape_res, Wei_scale, Sample_num=50):
# sample one
eps = torch.cuda.FloatTensor(Sample_num, Wei_shape_res.shape[0], Wei_shape_res.shape[1]).uniform_(0, 1)
theta = torch.unsqueeze(Wei_scale, axis=0).repeat(Sample_num, 1, 1) \
* torch.pow(-log_max(1 - eps), torch.unsqueeze(Wei_shape_res, axis=0).repeat(Sample_num, 1, 1)) #
theta = torch.max(theta, self.real_min.cuda())
return torch.mean(theta, dim=0, keepdim=False)
def compute_loss(self, x, re_x):
likelihood = torch.sum(x * self.log_max(re_x) - re_x - torch.lgamma(x + 1))
return - likelihood / x.shape[1]
def KL_GamWei(self, Gam_shape, Gam_scale, Wei_shape_res, Wei_scale):
eulergamma = torch.tensor(0.5772, dtype=torch.float32)
part1 = Gam_shape * self.log_max(Wei_scale) - eulergamma.cuda() * Gam_shape * Wei_shape_res + self.log_max(Wei_shape_res)
part2 = - Gam_scale * Wei_scale * torch.exp(torch.lgamma(1 + Wei_shape_res))
part3 = eulergamma.cuda() + 1 + Gam_shape * self.log_max(Gam_scale) - torch.lgamma(Gam_shape)
KL = part1 + part2 + part3
return - torch.sum(KL) / Wei_scale.shape[1]
def forward(self, x, train_flag=True):
hidden_list = [0] * self.layer_num
theta = [0] * self.layer_num
gam_scale = [0] * self.layer_num
k_rec = [0] * self.layer_num
l = [0] * self.layer_num
l_tmp = [0] * self.layer_num
phi_theta = [0] * self.layer_num
loss = [0] * (self.layer_num + 1)
likelihood = [0] * (self.layer_num + 1)
KL_dis = [0] * self.layer_num
graph_kl_loss = [0] * self.layer_num
throshold = [0] * self.layer_num
for t in range(self.layer_num):
if t == 0:
hidden = F.relu(self.bn_layer[t](self.h_encoder[t](x)))
else:
hidden = F.relu(self.bn_layer[t](self.h_encoder[t](hidden_list[t-1])))
hidden_list[t] = hidden
for t in range(self.layer_num-1, -1, -1):
if t == self.layer_num - 1:
k_rec_temp = torch.max(torch.nn.functional.softplus(self.shape_encoder[t](hidden_list[t])),
self.real_min.cuda()) # k_rec = 1/k
k_rec[t] = torch.min(k_rec_temp, self.wei_shape_max.cuda())
l_tmp[t] = torch.max(torch.nn.functional.softplus(self.scale_encoder[t](hidden_list[t])), self.real_min.cuda())
l[t] = l_tmp[t] / torch.exp(torch.lgamma(1 + k_rec[t]))
if train_flag:
if t == 0:
theta[t] = self.reparameterize(k_rec[t].permute(1, 0), l[t].permute(1, 0))
else:
theta[t] = self.reparameterize2(k_rec[t].permute(1, 0), l[t].permute(1, 0))
else:
theta[t] = l_tmp[t].permute(1, 0)
phi_theta[t], KL_dis[t] = self.decoder[t](theta[t], t)
else:
hidden_phitheta = torch.cat((hidden_list[t], phi_theta[t+1].permute(1, 0).detach()), 1)
k_rec_temp = torch.max(torch.nn.functional.softplus(self.shape_encoder[t](hidden_phitheta)),
self.real_min.cuda()) # k_rec = 1/k
k_rec[t] = torch.min(k_rec_temp, self.wei_shape_max.cuda())
l_tmp[t] = torch.max(torch.nn.functional.softplus(self.scale_encoder[t](hidden_phitheta)), self.real_min.cuda())
l[t] = l_tmp[t] / torch.exp(torch.lgamma(1 + k_rec[t]))
if train_flag:
if t == 0:
theta[t] = self.reparameterize(k_rec[t].permute(1, 0), l[t].permute(1, 0))
else:
theta[t] = self.reparameterize2(k_rec[t].permute(1, 0), l[t].permute(1, 0))
else:
theta[t] = l_tmp[t].permute(1, 0)
phi_theta[t], KL_dis[t] = self.decoder[t](theta[t], t)
for t in range(self.layer_num + 1):
if t == 0:
zero = torch.zeros_like(self.graph[t])
one = torch.ones_like(self.graph[t])
throshold[t] = torch.min(torch.where(self.graph[t] > 0, zero, one) * KL_dis[t], dim=0)[0] \
- torch.max(torch.where(self.graph[t] > 0, one, zero) * KL_dis[t], dim=0)[0]
graph_kl_loss[t] = 10*torch.mean(torch.relu(10.0 - throshold[t]))
likelihood[t] = self.compute_loss(x.permute(1, 0), phi_theta[t])
loss[t] = likelihood[t] + graph_kl_loss[t]
elif t == self.layer_num:
loss[t] = self.KL_GamWei(torch.tensor(1.0, dtype=torch.float32).cuda(), torch.tensor(1.0, dtype=torch.float32).cuda(),
k_rec[t - 1].permute(1, 0), l[t - 1].permute(1, 0))
likelihood[t] = loss[t]
else:
zero = torch.zeros_like(self.graph[t])
one = torch.ones_like(self.graph[t])
throshold[t] = torch.min(torch.where(self.graph[t] > 0, zero, one) * KL_dis[t], dim=0)[0] \
- torch.max(torch.where(self.graph[t] > 0, one, zero) * KL_dis[t], dim=0)[0]
graph_kl_loss[t] = 10*torch.mean(torch.relu(10.0 - throshold[t]))
likelihood[t] = self.KL_GamWei(phi_theta[t], torch.tensor(1.0, dtype=torch.float32).cuda(),
k_rec[t - 1].permute(1, 0), l[t - 1].permute(1, 0))
loss[t] = likelihood[t] + graph_kl_loss[t]
return phi_theta, theta, loss, likelihood, graph_kl_loss