-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
200 lines (163 loc) · 6.61 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
# -*- coding: utf-8 -*-
# ---
# @File: texture_mat.py
# @Author: sgdy3
# @E-mail: [email protected]
# @Time: 2022/3/22
# Describe: 实现了signet-f
# ---
from keras.layers import Conv2D,MaxPool2D,Dense,Input,BatchNormalization,Lambda,Activation,Flatten
from keras.models import Sequential,Model
from keras.losses import categorical_crossentropy,binary_crossentropy
import keras.backend as K
from keras.optimizer_v2.gradient_descent import SGD
import numpy as np
from keras.utils.vis_utils import plot_model
import os
class SigNet_F():
def __init__(self,num_class,mod='thin'):
self.rows=150
self.cols=220
self.channles=1
self.imgshape = (self.rows, self.cols, self.channles)
self.user_dim=num_class
self.batchsize=32
self.epochs=6
self.optimizer=SGD(lr=1e-3,momentum=0.9,nesterov=True,decay=5e-4)
assert mod=='thin' or 'std',"model has only two variant: thin and std"
if mod=='thin':
self.backbone=self.backbone_thin()
else:
self.backbone=self.backbone_std()
sig=Input(shape=self.imgshape)
m_label=Input(shape=(self.user_dim,))
f_label=Input(shape=(1,))
feature=self.backbone(sig)
pred_m=Dense(self.user_dim)(feature)
pred_f=Dense(1)(feature)
mixed_loss=Lambda(self.combine_loss,name='loss')([m_label,pred_m,f_label,pred_f])
self.signet_f=Model([sig,m_label,f_label],[pred_m,pred_f,mixed_loss])
loss_layer=self.signet_f.get_layer('loss').output
self.signet_f.add_loss(loss_layer)
self.signet_f.compile(optimizer=self.optimizer)
plot_model(self.signet_f, to_file='signet_f.png', show_shapes=True)
self.signet_f.summary()
def combine_loss(self,args,alpha=0.99):
m_label,pred_m,f_label,pred_f=args
cat_los=categorical_crossentropy(m_label,pred_m)
b_los=binary_crossentropy(f_label,pred_f)
return K.mean((1-alpha)*cat_los+alpha*b_los)
def backbone_thin(self):
seq=Sequential()
# 155*220->37*53
seq.add(Conv2D(32,kernel_size=11,strides=4,input_shape=self.imgshape,use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
# 37*53->18*26
seq.add(MaxPool2D(pool_size=3,strides=2))
# 18*26->8*12
seq.add(Conv2D(64,kernel_size=5,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
# 8*12->8*12
seq.add(MaxPool2D(pool_size=3,strides=2))
# 8*12->8*12
seq.add(Conv2D(64,kernel_size=3,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
# 8*12->8*12
seq.add(Conv2D(96,kernel_size=3,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
# 8*12->8*12
seq.add(Conv2D(96,kernel_size=5,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
# 8*12->3*5
seq.add(MaxPool2D(pool_size=3,strides=2))
# 3*5->2048*1
seq.add(Flatten())
seq.add(Dense(128,use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.summary()
# user_dim->binary
img=Input(shape=self.imgshape)
feature=seq(img)
return Model(img,feature)
def backbone_std(self):
seq=Sequential()
seq.add(Conv2D(96,kernel_size=11,strides=4,input_shape=self.imgshape,use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(MaxPool2D(pool_size=3,strides=2))
seq.add(Conv2D(256,kernel_size=5,strides=2,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(MaxPool2D(pool_size=3,strides=2))
seq.add(Conv2D(384,kernel_size=3,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(Conv2D(384,kernel_size=3,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(Conv2D(256,kernel_size=3,strides=1,padding='same',use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(MaxPool2D(pool_size=3,strides=2))
seq.add(Flatten())
seq.add(Dense(2048,use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.add(Dense(2048,use_bias=False))
seq.add(BatchNormalization())
seq.add(Activation('relu'))
seq.summary()
input=Input(shape=self.imgshape)
output=seq(input)
return Model(input,output)
def train(self,data,weights='',save=False):
save_dir = './NetWeights/Signet_f_weights'
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
if weights:
filepath = os.path.join(save_dir, weights)
self.signet_f.load_weights(filepath)
doc=None
else:
print('train')
filepath = os.path.join(save_dir, 'signet_f.h5')
train_img=data.shuffle(100).batch(self.batchsize).repeat(self.epochs)
time=0
doc=[]
for i in range(1,self.epochs+1):
for batch in train_img:
loss=self.signet_f.train_on_batch(batch)
doc.append(loss)
print("%d round: loss %f"%(time,loss))
time+=1
# 总共进行三次学习率下降，每次下降10%
if i%(self.epochs//3)==0:
self.optimizer.lr-=0.1*self.optimizer.lr
if save:
self.signet_f.save_weights(filepath)
return doc
def early_stop(stop_round,loss,pre_loss,threshold=0.005):
'''
early stop setting
:param stop_round: rounds under caculated
:param pre_loss: loss list
:param threshold: minimum one-order value of loss list
:return: whether or not to jump out
'''
if(len(pre_loss)<stop_round):
pre_loss.append(loss)
return False
else:
loss_diff=np.diff(pre_loss,1)
pre_loss.pop(0)
pre_loss.append(loss)
if(abs(loss_diff).mean()<threshold): # to low variance means flatten field
return True
else:
return False