-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbase_model.py
160 lines (135 loc) · 6.3 KB
/
base_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
import torch
import torch.nn as nn
from attention import Attention, NewAttention
from language_model import WordEmbedding, QuestionEmbedding, QuestionEmbedding2
from classifier import SimpleClassifier
from fc import FCNet
from Decoders.decoder1 import _netG as netG
import torch.nn.functional as F
from torch.autograd import Variable
from misc.utils import LayerNorm
class BaseModel2(nn.Module):
def __init__(self, w_emb, q_emb, h_emb, v_att, h_att, q_net, v_net, h_net, qih_att, qhi_att, qih_net, qhi_net,
decoder, args, qhih_att, qihi_att):
super(BaseModel2, self).__init__()
self.ninp = args.ninp
self.w_emb = w_emb
self.q_emb = q_emb
self.h_emb = h_emb
self.decoder = decoder
self.img_embed = nn.Linear(args.img_feat_size, 2 * args.nhid)
self.w1 = nn.Linear(args.nhid*2, args.nhid*2)
self.w2 = nn.Linear(args.nhid*2, args.nhid*2)
self.track_1 = v_att
self.locate_1 = h_att
self.locate_2 = qih_att
self.track_2 = qhi_att
self.locate_3 = qhih_att
self.track_3 = qihi_att
self.q_net = q_net
self.v_net = v_net
self.h_net = h_net
self.qih_net = qih_net
self.qhi_net = qhi_net
self.fc1 = nn.Linear(args.nhid * 4, self.ninp)
self.dropout = args.dropout
self.vocab_size = args.vocab_size
# self.fch = FCNet([args.nhid * 2, args.nhid * 2])
# self.layernorm = LayerNorm(args.nhid*2)
def forward(self, image, question, history, answer, tans, rnd, Training=True, sampling=False):
# prepare I, Q, H
image = self.img_embed(image)
w_emb = self.w_emb(question)
q_emb, ques_hidden = self.q_emb(w_emb) # [batch, q_dim]
hw_emb = self.w_emb(history)
h_emb, _ = self.h_emb(hw_emb) # [batch * rnd, h_dim]
h_emb = h_emb.view(-1, rnd, h_emb.size(1))
# cap & image
# qc_att = self.v_att(image, h_emb[:, 0, :])
# qc_emb = (qc_att * image).sum(1)
# qc_emb = self.fch(qc_emb * q_emb)
# question & image --> qi
qv_att = self.track_1(image, q_emb)
qv_emb = (qv_att * image).sum(1) # [batch, v_dim]
# question & history --> qh
qh_att = self.locate_1(h_emb, q_emb)
qh_emb = (qh_att * h_emb).sum(1) # [batch, h_dim]
# qh_emb = self.fch(qh_emb+q_emb)
# qh_emb = self.layernorm(qh_emb+h_emb[:,0,:])
# qh & image --> qhi
qhi_att = self.track_2(image, qh_emb)
qhi_emb = (qhi_att * image).sum(1) # [batch, v_dim]
# qi & history --> qih
qih_att = self.locate_2(h_emb, qv_emb)
qih_emb = (qih_att * h_emb).sum(1) # [batch, h_dim]
q_re = self.q_net(q_emb)
qih_emb = self.h_net(qih_emb)
qih_emb = q_re * qih_emb
qhi_emb = self.v_net(qhi_emb)
qhi_emb = q_re * qhi_emb
# qih & i --> qihi
qihi_att = self.track_3(image, qih_emb)
qihi_emb = (qihi_att * image).sum(1)
# qhi & his --> qhih
qhih_att = self.locate_3(h_emb, qhi_emb)
qhih_emb = (qhih_att * h_emb).sum(1)
q_repr = self.q_net(q_emb)
qhi_repr = self.qhi_net(qihi_emb)
qqhi_joint_repr = q_repr * qhi_repr
qih_repr = self.qih_net(qhih_emb)
qqih_joint_repr = q_repr * qih_repr
joint_repr = torch.cat([self.w1(qqhi_joint_repr), self.w2(qqih_joint_repr)], 1) # [batch, h_dim * 2
joint_repr = F.tanh(self.fc1(F.dropout(joint_repr, self.dropout, training=self.training)))
_, ques_hidden = self.decoder(joint_repr.view(-1, 1, self.ninp), ques_hidden)
if sampling:
batch_size, _, _ = image.size()
sample_ans_input = Variable(torch.LongTensor(batch_size, 1).fill_(2).cuda())
sample_opt = {'beam_size': 1}
seq, seqLogprobs = self.decoder.sample(self.w_emb, sample_ans_input, ques_hidden, sample_opt)
sample_ans = self.w_emb(Variable(seq))
ans_emb = self.w_emb(tans)
sample_ans = torch.cat([w_emb, joint_repr.view(batch_size, -1, self.ninp),sample_ans], 1)
ans_emb = torch.cat([w_emb, joint_repr.view(batch_size, -1, self.ninp), ans_emb], 1)
return sample_ans, ans_emb
if not Training:
batch_size, _, hid_size = image.size()
hid_size = int(hid_size / 2)
hidden_replicated = []
for hid in ques_hidden:
hidden_replicated.append(hid.view(2, batch_size, 1,hid_size).expand(2,
batch_size, 100, hid_size).clone().view(2, -1, hid_size))
hidden_replicated = tuple(hidden_replicated)
ques_hidden = hidden_replicated
emb = self.w_emb(answer)
pred, _ = self.decoder(emb, ques_hidden)
return pred
def build_baseline0_newatt2(args, num_hid):
w_emb = WordEmbedding(args.vocab_size, args.ninp, 0.0)
q_emb = QuestionEmbedding2(args.ninp, num_hid, args.nlayers, True, 0.0)
h_emb = QuestionEmbedding2(args.ninp, num_hid, args.nlayers, True, 0.0)
v_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
h_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
qih_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
qhi_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
q_net = FCNet([q_emb.num_hid*2, num_hid*2])
v_net = FCNet([args.nhid*2, num_hid*2])
h_net = FCNet([args.nhid*2, num_hid*2])
qih_net = FCNet([args.nhid*2, num_hid*2])
qhi_net = FCNet([args.nhid*2, num_hid*2])
qhih_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
qihi_att = NewAttention(args.nhid*2, q_emb.num_hid*2, num_hid*2)
decoder = netG(args)
return BaseModel2(w_emb, q_emb, h_emb, v_att, h_att, q_net, v_net, h_net, qih_att, qhi_att, qih_net, qhi_net,
decoder, args, qhih_att, qihi_att)
class attflat(nn.Module):
def __init__(self, args):
super(attflat, self).__init__()
self.mlp = FCNet([args.nhid * 2, args.nhid, 1])
self.fc = nn.Linear(args.nhid*2, args.nhid*2)
def forward(self, x):
batch_size, q_len, nhid = x.size()
att = self.mlp(x.view(-1, nhid))
att = F.softmax(att, dim=1)
x_atted = (att.view(batch_size, q_len, -1) * x.view(batch_size, q_len, -1)).sum(1)
x_atted = self.fc(x_atted)
return x_atted