forked from laoreja/face-identity-transformer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
net_sphere.py
199 lines (161 loc) · 6.96 KB
/
net_sphere.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import Parameter
import math
def myphi(x, m):
x = x * m
return 1 - x ** 2 / math.factorial(2) + x ** 4 / math.factorial(4) - x ** 6 / math.factorial(6) + \
x ** 8 / math.factorial(8) - x ** 9 / math.factorial(9)
class AngleLinear(nn.Module):
def __init__(self, in_features, out_features, m=4, phiflag=True):
super(AngleLinear, self).__init__()
self.in_features = in_features
self.out_features = out_features
# weight_data = torch.rand((in_features, out_features)) * 2 - 1.
# weight_data.renorm_(2, 1, 1e-5).mul_(1e5)
# self.weight = Parameter(weight_data)
self.weight = Parameter(torch.Tensor(in_features, out_features))
self.weight.data.uniform_(-1, 1).renorm_(2, 1, 1e-5).mul_(1e5)
self.phiflag = phiflag
self.m = m
self.mlambda = [
lambda x: x ** 0,
lambda x: x ** 1,
lambda x: 2 * x ** 2 - 1,
lambda x: 4 * x ** 3 - 3 * x,
lambda x: 8 * x ** 4 - 8 * x ** 2 +
1,
lambda x: 16 * x ** 5 - 20 * x ** 3 + 5 * x
]
def forward(self, input):
x = input # size=(B,F) F is feature len
w = self.weight # size=(F,Classnum) F=in_features Classnum=out_features
ww = w.renorm(2, 1, 1e-5).mul(1e5)
xlen = x.pow(2).sum(1).pow(0.5) # size=B
wlen = ww.pow(2).sum(0).pow(0.5) # size=Classnum
cos_theta = x.mm(ww) # size=(B,Classnum)
cos_theta = cos_theta / xlen.view(-1, 1) / wlen.view(1, -1)
cos_theta = cos_theta.clamp(-1, 1)
if self.phiflag:
cos_m_theta = self.mlambda[self.m](cos_theta)
theta = cos_theta.detach().acos()
k = (self.m * theta / 3.14159265).floor()
n_one = - 1
phi_theta = (n_one ** k) * cos_m_theta - 2 * k
else:
theta = cos_theta.acos()
phi_theta = myphi(theta, self.m)
phi_theta = phi_theta.clamp(-1 * self.m, 1)
cos_theta = cos_theta * xlen.view(-1, 1)
phi_theta = phi_theta * xlen.view(-1, 1)
output = (cos_theta, phi_theta)
return output # size=(B,Classnum,2)
class AngleLoss(nn.Module):
def __init__(self, gamma=0):
super(AngleLoss, self).__init__()
self.gamma = gamma
self.it = 0
self.LambdaMin = 5.0
self.LambdaMax = 1500.0
self.lamb = 1500.0
def forward(self, input, target):
self.it += 1
cos_theta, phi_theta = input
target = target.view(-1, 1) # size=(B,1)
index = torch.zeros_like(cos_theta) # size=(B,Classnum)
index.scatter_(1, target.detach(), 1)
index = index.byte()
self.lamb = max(self.LambdaMin, self.LambdaMax / (1 + 0.1 * self.it))
output = cos_theta * 1.0 # size=(B,Classnum)
output[index] -= cos_theta[index] * (1.0 + 0) / (1 + self.lamb)
output[index] += phi_theta[index] * (1.0 + 0) / (1 + self.lamb)
logpt = F.log_softmax(output)
logpt = logpt.gather(1, target)
logpt = logpt.view(-1)
pt = logpt.detach().exp()
loss = -1 * (1 - pt) ** self.gamma * logpt
loss = loss.mean()
return loss
class sphere20a(nn.Module):
def __init__(self, feature, classnum=10574):
# input needs to be aligned
# by default, classnum=10574
# during Train, feature = False; During evaluation, feature=True
super(sphere20a, self).__init__()
self.classnum = classnum
# if feature is True:
# feature = 5
self.feature = feature
# input = B*3*112*96
self.conv1_1 = nn.Conv2d(3, 64, 3, 2, 1) # =>B*64*56*48
self.relu1_1 = nn.PReLU(64)
self.conv1_2 = nn.Conv2d(64, 64, 3, 1, 1)
self.relu1_2 = nn.PReLU(64)
self.conv1_3 = nn.Conv2d(64, 64, 3, 1, 1)
self.relu1_3 = nn.PReLU(64)
self.conv2_1 = nn.Conv2d(64, 128, 3, 2, 1) # =>B*128*28*24
self.relu2_1 = nn.PReLU(128)
self.conv2_2 = nn.Conv2d(128, 128, 3, 1, 1)
self.relu2_2 = nn.PReLU(128)
self.conv2_3 = nn.Conv2d(128, 128, 3, 1, 1)
self.relu2_3 = nn.PReLU(128)
self.conv2_4 = nn.Conv2d(128, 128, 3, 1, 1) # =>B*128*28*24
self.relu2_4 = nn.PReLU(128)
self.conv2_5 = nn.Conv2d(128, 128, 3, 1, 1)
self.relu2_5 = nn.PReLU(128)
self.conv3_1 = nn.Conv2d(128, 256, 3, 2, 1) # =>B*256*14*12
self.relu3_1 = nn.PReLU(256)
self.conv3_2 = nn.Conv2d(256, 256, 3, 1, 1)
self.relu3_2 = nn.PReLU(256)
self.conv3_3 = nn.Conv2d(256, 256, 3, 1, 1)
self.relu3_3 = nn.PReLU(256)
self.conv3_4 = nn.Conv2d(256, 256, 3, 1, 1) # =>B*256*14*12
self.relu3_4 = nn.PReLU(256)
self.conv3_5 = nn.Conv2d(256, 256, 3, 1, 1)
self.relu3_5 = nn.PReLU(256)
self.conv3_6 = nn.Conv2d(256, 256, 3, 1, 1) # =>B*256*14*12
self.relu3_6 = nn.PReLU(256)
self.conv3_7 = nn.Conv2d(256, 256, 3, 1, 1)
self.relu3_7 = nn.PReLU(256)
self.conv3_8 = nn.Conv2d(256, 256, 3, 1, 1) # =>B*256*14*12
self.relu3_8 = nn.PReLU(256)
self.conv3_9 = nn.Conv2d(256, 256, 3, 1, 1)
self.relu3_9 = nn.PReLU(256)
self.conv4_1 = nn.Conv2d(256, 512, 3, 2, 1) # =>B*512*7*6
self.relu4_1 = nn.PReLU(512)
self.conv4_2 = nn.Conv2d(512, 512, 3, 1, 1)
self.relu4_2 = nn.PReLU(512)
self.conv4_3 = nn.Conv2d(512, 512, 3, 1, 1)
self.relu4_3 = nn.PReLU(512)
self.fc5 = nn.Linear(512 * 7 * 6, 512)
self.fc6 = AngleLinear(512, self.classnum)
def forward(self, x):
x = self.relu1_1(self.conv1_1(x))
x = x + self.relu1_3(self.conv1_3(self.relu1_2(self.conv1_2(x))))
if self.feature == 1:
res_feature = x.view(x.size(0), -1)
x = self.relu2_1(self.conv2_1(x))
x = x + self.relu2_3(self.conv2_3(self.relu2_2(self.conv2_2(x))))
x = x + self.relu2_5(self.conv2_5(self.relu2_4(self.conv2_4(x))))
if self.feature == 2:
res_feature = x.view(x.size(0), -1)
x = self.relu3_1(self.conv3_1(x))
x = x + self.relu3_3(self.conv3_3(self.relu3_2(self.conv3_2(x))))
x = x + self.relu3_5(self.conv3_5(self.relu3_4(self.conv3_4(x))))
x = x + self.relu3_7(self.conv3_7(self.relu3_6(self.conv3_6(x))))
x = x + self.relu3_9(self.conv3_9(self.relu3_8(self.conv3_8(x))))
if self.feature == 3:
res_feature = x.view(x.size(0), -1)
x = self.relu4_1(self.conv4_1(x))
x = x + self.relu4_3(self.conv4_3(self.relu4_2(self.conv4_2(x))))
if self.feature == 4:
res_feature = x.view(x.size(0), -1)
x = x.view(x.size(0), -1)
x = self.fc5(x)
if self.feature == 5 or self.feature is True:
res_feature = x
if self.feature is True:
return res_feature
x = self.fc6(x)
return x, res_feature