-
Notifications
You must be signed in to change notification settings - Fork 18
/
model.py
207 lines (142 loc) · 6.23 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
#encoding:utf-8
# -----------------------------------------------------------
# "Matching Images and Text with Multi-modal Tensor Fusion and Re-ranking"
# WangTan, XingXu, YangYang, Alan Hanjalic, HengtaoShen, JingkuanSong
# ACM Multimedia 2019, Nice, France
# Writen by WangTan, 2019
# ------------------------------------------------------------
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import copy
from utils import collect_match, collect_neg, acc_train
import seq2vec
class AbstractNoAtt(nn.Module):
def __init__(self, opt={}, vocab_words=[]):
super(AbstractNoAtt, self).__init__()
self.opt = opt
self.vocab_words = vocab_words
self.num_classes = 1
# Modules
self.seq2vec = seq2vec.factory(self.vocab_words, self.opt['seq2vec'])
# Modules for classification
self.linear_classif = nn.Linear(self.opt['fusion']['dim_h'], self.num_classes)
self.Eiters = 0
def _fusion(self, input_v, input_q):
raise NotImplementedError
def _classif(self, x):
batch_size_v = x.size(0)
batch_size_t = x.size(1)
if 'activation' in self.opt['classif']:
x = getattr(F, self.opt['classif']['activation'])(x)
x = F.dropout(x,
p=self.opt['classif']['dropout'],
training=self.training)
x = self.linear_classif(x)
x = torch.sigmoid(x)
x = x.view( batch_size_v, batch_size_t)
return x
def forward(self, input_v, input_t):
if input_v.dim() != 4 and input_t.dim() != 2:
raise ValueError
self.Eiters += 1
batch_size = input_v.size(0)
x_t_vec = self.seq2vec(input_t)
x_v = torch.mean(input_v, 1)
x = self._fusion(x_v, x_t_vec)
x = self._classif(x)
# """calculate acc during training"""
# if self.training:
# acc, recall, precision = acc_train(x_m.cpu().clone().data)
# self.logger.update('acc', acc)
# self.logger.update('rec', recall)
# self.logger.update('pre', precision)
return x
class FusionNoAtt(AbstractNoAtt):
def __init__(self, opt={}, vocab_words=[]):
opt['fusion']['dim_h'] = opt['fusion']['dim_mm']
super(FusionNoAtt, self).__init__(opt, vocab_words)
# Modules for classification
self.fusion = Core_Fusion(self.opt['fusion'])
def _fusion(self, x_v, x_t):
return self.fusion(x_v, x_t)
class AbstractFusion(nn.Module):
def __init__(self, opt={}):
super(AbstractFusion, self).__init__()
self.opt = opt
def forward(self, input_v, input_t):
raise NotImplementedError('input should be visual and language')
class Core_Fusion(AbstractFusion):
def __init__(self, opt):
super(Core_Fusion, self).__init__(opt)
# visul & text embedding
self.linear_v = nn.Linear(self.opt['dim_v'], self.opt['dim_hv'])
self.linear_t = nn.Linear(self.opt['dim_t'], self.opt['dim_ht'])
# Core tensor
self.list_linear_hv = nn.ModuleList([
nn.Linear(self.opt['dim_hv'], self.opt['dim_mm'])
for i in range(self.opt['R'])])
self.list_linear_ht = nn.ModuleList([
nn.Linear(self.opt['dim_ht'], self.opt['dim_mm'])
for i in range(self.opt['R'])])
def forward(self, input_v, input_t):
if input_v.dim() != input_t.dim() and input_v.dim() != 3:
raise ValueError
batch_size_v = input_v.size(0)
batch_size_t = input_t.size(0)
x_v = F.dropout(input_v, p=self.opt['dropout_v'], training=self.training)
x_v = self.linear_v(x_v)
if 'activation_v' in self.opt:
x_v = getattr(F, self.opt['activation_v'])(x_v)
x_t = F.dropout(input_t, p=self.opt['dropout_t'], training=self.training)
x_t = self.linear_t(x_t)
if 'activation_t' in self.opt:
x_t = getattr(F, self.opt['activation_t'])(x_t)
x_mm = []
for i in range(self.opt['R']):
x_hv = F.dropout(x_v, p=self.opt['dropout_hv'], training=self.training)
x_hv = self.list_linear_hv[i](x_hv)
if 'activation_hv' in self.opt:
x_hv = getattr(F, self.opt['activation_hv'])(x_hv)
x_ht = F.dropout(x_t, p=self.opt['dropout_ht'], training=self.training)
x_ht = self.list_linear_ht[i](x_ht)
if 'activation_ht' in self.opt:
x_ht = getattr(F, self.opt['activation_ht'])(x_ht)
x_mm.append(torch.mul(x_hv[:, None, :], x_ht[None, :, :]))
x_mm = torch.stack(x_mm, dim=1)
x_mm = x_mm.sum(1).view(batch_size_v, batch_size_t, self.opt['dim_mm'])
if 'activation_mm' in self.opt:
x_mm = getattr(F, self.opt['activation_mm'])(x_mm)
return x_mm
class Fusion2d(Core_Fusion):
def __init__(self, opt):
super(Fusion2d, self).__init__(opt)
def forward(self, input_v, input_t):
if input_v.dim() != input_t.dim() and input_v.dim() != 3:
raise ValueError
batch_size = input_v.size(0)
weight_height = input_v.size(1)
dim_hv = input_v.size(2)
dim_ht = input_t.size(2)
if not input_v.is_contiguous():
input_v = input_v.contiguous()
if not input_t.is_contiguous():
input_t = input_t.contiguous()
x_v = input_v.view(batch_size * weight_height, self.opt['dim_hv'])
x_t = input_t.view(batch_size * weight_height, self.opt['dim_ht'])
x_mm = super().forward(x_v, x_t)
if not x_mm.is_contiguous():
x_mm = x_mm.contiguous()
x_mm = x_mm.view(batch_size, batch_size, weight_height, self.opt['dim_mm'])
return x_mm
def factory(opt, vocab_words, cuda=True, data_parallel=True):
opt = copy.copy(opt)
model = FusionNoAtt(opt, vocab_words)
if data_parallel:
model = nn.DataParallel(model).cuda()
if not cuda:
raise ValueError
if cuda:
model.cuda()
return model