-
Notifications
You must be signed in to change notification settings - Fork 2
/
model.py
186 lines (155 loc) · 8.49 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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from utils import process_super_class
class BayesianHead(nn.Module):
"""
The prediction head with a hierarchical classification when the optional transformer encoder is used.
"""
def __init__(self, input_dim=512, num_geometric=15, num_possessive=11, num_semantic=24, T1=1, T2=1, T3=1):
super(BayesianHead, self).__init__()
self.fc3_1 = nn.Linear(input_dim, num_geometric)
self.fc3_2 = nn.Linear(input_dim, num_possessive)
self.fc3_3 = nn.Linear(input_dim, num_semantic)
self.fc5 = nn.Linear(input_dim, 3)
self.T1 = T1
self.T2 = T2
self.T3 = T3
def forward(self, h):
super_relation = F.log_softmax(self.fc5(h), dim=1)
# By Bayes rule, log p(relation_n, super_n) = log p(relation_1 | super_1) + log p(super_1)
relation_1 = self.fc3_1(h) # geometric
relation_1 = F.log_softmax(relation_1 / self.T1, dim=1) + super_relation[:, 0].view(-1, 1)
relation_2 = self.fc3_2(h) # possessive
relation_2 = F.log_softmax(relation_2 / self.T2, dim=1) + super_relation[:, 1].view(-1, 1)
relation_3 = self.fc3_3(h) # semantic
relation_3 = F.log_softmax(relation_3 / self.T3, dim=1) + super_relation[:, 2].view(-1, 1)
return relation_1, relation_2, relation_3, super_relation
class FlatRelationClassifier(nn.Module):
"""
The local prediction module with a flat classification.
"""
def __init__(self, args, input_dim=128, output_dim=50, feature_size=32, num_classes=150, num_super_classes=17):
super(FlatRelationClassifier, self).__init__()
self.num_classes = num_classes
self.num_super_classes = num_super_classes
self.conv1_1 = nn.Conv2d(2 * input_dim + 1, input_dim, kernel_size=1, stride=1, padding=0)
self.conv1_2 = nn.Conv2d(2 * input_dim + 1, input_dim, kernel_size=1, stride=1, padding=0)
self.conv2_1 = nn.Conv2d(2 * input_dim, 4 * input_dim, kernel_size=3, stride=1, padding=1)
self.conv3_1 = nn.Conv2d(4 * input_dim, 8 * input_dim, kernel_size=3, stride=1, padding=1)
self.dropout1 = nn.Dropout(p=0.5)
self.dropout2 = nn.Dropout(p=0.5)
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(8 * input_dim * (feature_size // 4) ** 2, 4096)
if args['dataset']['dataset'] == 'vg':
self.fc2 = nn.Linear(4096 + 2 * (num_classes + num_super_classes), 512)
else:
self.fc2 = nn.Linear(4096 + 2 * num_classes, 512)
self.fc3 = nn.Linear(512, output_dim)
self.fc4 = nn.Linear(512, 1)
def conv_layers(self, h_sub, h_obj):
h_sub = torch.tanh(self.conv1_1(h_sub))
h_obj = torch.tanh(self.conv1_2(h_obj))
h = torch.cat((h_sub, h_obj), dim=1) # (batch_size, 256, 32, 32)
h = F.relu(self.conv2_1(h)) # (batch_size, 512, 32, 32)
h = self.maxpool(h) # (batch_size, 512, 16, 16)
h = F.relu(self.conv3_1(h)) # (batch_size, 1024,16, 16)
h = self.maxpool(h) # (batch_size, 1024, 8, 8)
h = torch.reshape(h, (h.shape[0], -1))
h = self.dropout1(F.relu(self.fc1(h)))
return h
def concat_labels(self, h, c1, c2, s1, s2, rank, h_aug=None):
c1 = F.one_hot(c1, num_classes=self.num_classes)
c2 = F.one_hot(c2, num_classes=self.num_classes)
if s1 is not None: # concatenate super-class labels as well
s1, s2 = process_super_class(s1, s2, self.num_super_classes, rank)
hc = torch.cat((h, c1, c2, s1, s2), dim=1)
if h_aug is not None:
h_aug = torch.cat((h_aug, c1, c2, s1, s2), dim=1)
h_aug = self.dropout2(F.relu(self.fc2(h_aug)))
else:
hc = torch.cat((h, c1, c2), dim=1)
if h_aug is not None:
h_aug = torch.cat((h_aug, c1, c2), dim=1)
h_aug = self.dropout2(F.relu(self.fc2(h_aug)))
return hc, h_aug
def forward(self, h_sub, h_obj, c1, c2, s1, s2, rank, h_sub_aug=None, h_obj_aug=None, one_hot=True):
h = self.conv_layers(h_sub, h_obj)
h_aug = self.conv_layers(h_sub_aug, h_obj_aug) if h_sub_aug is not None else None # need data augmentation in contrastive learning
hc, pred_aug = self.concat_labels(h, c1, c2, s1, s2, rank, h_aug)
pred = self.dropout2(F.relu(self.fc2(hc)))
relation = self.fc3(pred) # (batch_size, 50)
connectivity = self.fc4(pred) # (batch_size, 1)
return relation, connectivity, pred, pred_aug
class BayesianRelationClassifier(nn.Module):
"""
The local prediction module with a hierarchical classification.
"""
def __init__(self, args, input_dim=128, feature_size=32, num_classes=150, num_super_classes=17, num_geometric=15,
num_possessive=11, num_semantic=24, T1=1, T2=1, T3=1):
super(BayesianRelationClassifier, self).__init__()
self.input_dim = input_dim
self.num_classes = num_classes
self.num_super_classes = num_super_classes
self.conv1_1 = nn.Conv2d(2 * input_dim + 1, input_dim, kernel_size=1, stride=1, padding=0)
self.conv1_2 = nn.Conv2d(2 * input_dim + 1, input_dim, kernel_size=1, stride=1, padding=0)
self.conv2_1 = nn.Conv2d(2 * input_dim, 4 * input_dim, kernel_size=3, stride=1, padding=1)
self.conv3_1 = nn.Conv2d(4 * input_dim, 8 * input_dim, kernel_size=3, stride=1, padding=1)
self.dropout1 = nn.Dropout(p=0.5)
self.dropout2 = nn.Dropout(p=0.5)
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(8 * input_dim * (feature_size // 4) ** 2, 4096)
if args['dataset']['dataset'] == 'vg':
self.fc2 = nn.Linear(4096 + 2 * (num_classes + num_super_classes), 512)
else:
self.fc2 = nn.Linear(4096 + 2 * num_classes, 512)
self.fc3_1 = nn.Linear(512, num_geometric)
self.fc3_2 = nn.Linear(512, num_possessive)
self.fc3_3 = nn.Linear(512, num_semantic)
self.fc4 = nn.Linear(512, 1)
self.fc5 = nn.Linear(512, 3)
self.T1 = T1
self.T2 = T2
self.T3 = T3
def conv_layers(self, h_sub, h_obj):
h_sub = torch.tanh(self.conv1_1(h_sub))
h_obj = torch.tanh(self.conv1_2(h_obj))
h = torch.cat((h_sub, h_obj), dim=1) # (batch_size, 256, 32, 32)
h = F.relu(self.conv2_1(h)) # (batch_size, 512, 32, 32)
h = self.maxpool(h) # (batch_size, 512, 16, 16)
h = F.relu(self.conv3_1(h)) # (batch_size, 1024,16, 16)
h = self.maxpool(h) # (batch_size, 1024, 8, 8)
h = torch.reshape(h, (h.shape[0], -1))
h = self.dropout1(F.relu(self.fc1(h)))
return h
def concat_labels(self, h, c1, c2, s1, s2, rank, h_aug=None):
c1 = F.one_hot(c1, num_classes=self.num_classes)
c2 = F.one_hot(c2, num_classes=self.num_classes)
if s1 is not None: # concatenate super-class labels as well
s1, s2 = process_super_class(s1, s2, self.num_super_classes, rank)
hc = torch.cat((h, c1, c2, s1, s2), dim=1)
if h_aug is not None:
h_aug = torch.cat((h_aug, c1, c2, s1, s2), dim=1)
h_aug = self.dropout2(F.relu(self.fc2(h_aug)))
else:
hc = torch.cat((h, c1, c2), dim=1)
if h_aug is not None:
h_aug = torch.cat((h_aug, c1, c2), dim=1)
h_aug = self.dropout2(F.relu(self.fc2(h_aug)))
return hc, h_aug
def forward(self, h_sub, h_obj, c1, c2, s1, s2, rank, h_sub_aug=None, h_obj_aug=None):
h = self.conv_layers(h_sub, h_obj)
h_aug = self.conv_layers(h_sub_aug, h_obj_aug) if h_sub_aug is not None else None # need data augmentation in contrastive learning
hc, pred_aug = self.concat_labels(h, c1, c2, s1, s2, rank, h_aug)
pred = self.dropout2(F.relu(self.fc2(hc)))
connectivity = self.fc4(pred) # (batch_size, 1)
super_relation = F.log_softmax(self.fc5(pred), dim=1)
relation_1 = self.fc3_1(pred) # geometric
relation_1 = F.log_softmax(relation_1 / self.T1, dim=1) + super_relation[:, 0].view(-1, 1)
relation_2 = self.fc3_2(pred) # possessive
relation_2 = F.log_softmax(relation_2 / self.T2, dim=1) + super_relation[:, 1].view(-1, 1)
relation_3 = self.fc3_3(pred) # semantic
relation_3 = F.log_softmax(relation_3 / self.T3, dim=1) + super_relation[:, 2].view(-1, 1)
return relation_1, relation_2, relation_3, super_relation, connectivity, pred, pred_aug