-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
167 lines (132 loc) · 5.45 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
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
class ConvInput(nn.Module):
'''Convolution Layers for Visual Inputs'''
def __init__(self):
super(ConvInput, self).__init__()
self.conv1 = nn.Conv2d(3, 24, 3, stride=2, padding=1)
self.batchNorm1 = nn.BatchNorm2d(24)
self.conv2 = nn.Conv2d(24, 24, 3, stride=2, padding=1)
self.batchNorm2 = nn.BatchNorm2d(24)
self.conv3 = nn.Conv2d(24, 24, 3, stride=2, padding=1)
self.batchNorm3 = nn.BatchNorm2d(24)
self.conv4 = nn.Conv2d(24, 24, 3, stride=2, padding=1)
self.batchNorm4 = nn.BatchNorm2d(24)
def forward(self, img):
x = self.conv1(img)
x = F.relu(x)
x = self.batchNorm1(x)
x = self.conv2(x)
x = F.relu(x)
x = self.batchNorm2(x)
x = self.conv3(x)
x = F.relu(x)
x = self.batchNorm3(x)
x = self.conv4(x)
x = F.relu(x)
x = self.batchNorm4(x)
return x
class BaseLayer(nn.Module):
'''Base Layer'''
def __init__(self):
super(BaseLayer, self).__init__()
def train_(self, img, ques, ans):
'''Train on a batch of train data'''
self.optimizer.zero_grad()
# forward pass
output = self(img, ques)
# loss and optimization
loss = F.nll_loss(output, ans)
loss.backward()
self.optimizer.step()
# accuracy
pred = output.argmax(1)
accuracy = pred.eq(ans.data).cpu().sum() * 100. / len(ans)
return accuracy, loss
def evaluate(self, img, ques, ans):
'''Evaluate on a batch of test data'''
# forward pass
output = self(img, ques)
# loss
loss = F.nll_loss(output, ans)
# accuracy
pred = output.argmax(1)
accuracy = pred.eq(ans.data).cpu().sum() * 100. / len(ans)
return accuracy, loss
def save_model(self, epoch):
torch.save(self.state_dict(), 'models/epoch_{:02d}.pth'.format(epoch))
class RNLayer(nn.Module):
'''Relation Network Layer (g_θ)'''
def __init__(self):
super(RNLayer, self).__init__()
self.g_fc1 = nn.Linear(62, 256) # 62 i.e. (24+2)*2+10
self.g_fc2 = nn.Linear(256, 256)
self.g_fc3 = nn.Linear(256, 256)
self.g_fc4 = nn.Linear(256, 256)
def forward(self, x, ques):
'''Forward Pass through g_θ'''
# x -> (b, 24, 5, 5)
# que -> (b, 10)
b, c, d, _ = x.size()
# add coordinates to x (feature maps)
self.build_coord_tensor(b, d) # coord_tensor -> (b, 2, 5, 5)
if x.device.type == 'cuda':
self.coord_tensor = self.coord_tensor.cuda()
x_tagged = torch.cat([x, self.coord_tensor], 1) # (b, 24+2, 5, 5)
x_flat = x_tagged.view(b, c+2, d*d).permute(0, 2, 1) # (b, 25, 24+2)
# repeat question vector for casting everywhere
ques = ques.unsqueeze(1) # (b, 1, 10)
ques = ques.repeat(1, 25, 1) # (b, 25, 10)
ques = ques.unsqueeze(2) # (b, 25, 1, 10)
# create i-j pairs of objects
x_i = x_flat.unsqueeze(1) # (b, 1, 25, 26)
x_i = x_i.repeat(1, 25, 1, 1) # (b, 25, 25, 26)
x_j = x_flat.unsqueeze(2) # (b, 25, 1, 26)
# add question vector
x_j = torch.cat([x_j, ques], 3) # (b, 25, 1, 26+10)
x_j = x_j.repeat(1, 1, 25, 1) # (b, 25, 25, 36)
# cast the pairs together
x_full = torch.cat([x_i, x_j], 3) # (b, 25, 25, 26*2+10)
# flatten for passing through g_θ
x_g = x_full.view(-1, 62) # (b * 25 * 25, 62)
# forward pass through g_θ linear layers
x_g = F.relu(self.g_fc1(x_g))
x_g = F.relu(self.g_fc2(x_g))
x_g = F.relu(self.g_fc3(x_g))
x_g = F.relu(self.g_fc4(x_g))
# reshape
x_g = x_g.view(b, (d * d) * (d * d), 256) # (b, 625, 256)
# sum
x_g = x_g.sum(1).squeeze(1) # (b, 256)
return x_g
def build_coord_tensor(self, b, d):
'''Returns the coordinates of the objects (d = 5)'''
coords = torch.linspace(-d/2., d/2., d) # (5)
x = coords.unsqueeze(0).repeat(d, 1) # (5, 5)
y = coords.unsqueeze(1).repeat(1, d) # (5, 5)
ct = torch.stack((x,y)) # (2, 5, 5)
self.coord_tensor = ct.unsqueeze(0).repeat(b, 1, 1, 1) # (b, 2, 5, 5)
class RNModel(BaseLayer):
'''Relation Network Model (CNN + g_θ + f_ϕ)'''
def __init__(self, args):
super(RNModel, self).__init__()
l_rate = args.lr if args != None else 0.0001
self.conv_input = ConvInput() # CNN features
self.rel_layer = RNLayer() # g_θ
# f_ϕ
self.f_fc1 = nn.Linear(256, 256)
self.f_fc2 = nn.Linear(256, 256)
self.f_out = nn.Linear(256, 10) # outputs logits over answer vocabulary
# Define the optimizer
self.optimizer = torch.optim.Adam(self.parameters(), lr=l_rate)
def forward(self, img, ques):
'''Forward pass through the full Relation Network Augmented model'''
x = self.conv_input(img) # (b, 24, 5, 5)
x_g = self.rel_layer(x, ques)
x_f = F.relu(self.f_fc1(x_g))
x_f = F.relu(self.f_fc2(x_f))
x_f = F.dropout(x_f)
out = F.log_softmax(self.f_out(x_f), dim=1)
return out