-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathunet.py
164 lines (136 loc) · 6.5 KB
/
unet.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
from keras.models import Model
from keras.layers import Input, concatenate, MaxPooling2D,Conv2D, Activation, UpSampling2D, BatchNormalization
from keras.optimizers import RMSprop, Adadelta, SGD, Adam
from keras.applications.vgg19 import VGG19
from keras.applications.vgg16 import VGG16
from unet.coord import *
from unet.losses import *
import numpy as np
from keras_applications.resnet import ResNet101
import keras
import keras_applications
keras_applications.set_keras_submodules(
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils
)
class UNet():
def __init__(self):
self.model = None
self.batch_norm = None
self.encoder_type = None
def get_model(self):
return self.model
def create_model(self,batch_norm=True,input_shape=(512,768,3),feature_maps=[16,32,64,128,256],num_classes=2):
self.batch_norm = batch_norm
concats_list = []
input = Input(shape=input_shape)
origin = input
# downsampling:
i=0
for f in feature_maps:
i+=1
block_name = 'encoder_block' + str(i) + "_conv"
output, concat_layer = self.__encoder_block(input,f,block_name)
input = output
concats_list.append(concat_layer)
# center
center = input
for _ in range(2):
center = Conv2D(feature_maps[-1]*2, (3, 3), padding='same')(center)
center = BatchNormalization()(center)
center = Activation('relu')(center)
input = center
# upsampling:
i=0
for f,c in zip(feature_maps[::-1],concats_list[::-1]):
i+=1
block_name = 'decoder_block' + str(i) + "_conv"
output = self.__decoder_block(input,c,f,block_name)
input = output
final_layer = Conv2D(num_classes, (3, 3), padding='same')(output)
final_layer = Activation("sigmoid")(final_layer)
self.model = Model(inputs=origin, outputs=final_layer)
self.__compile_model
def create_pretrained_model(self,encoder_type='vgg19',batch_norm=False,coord_conv=False,input_shape=(512, 768, 3),num_classes=2):
self.encoder_type = encoder_type
self.batch_norm = batch_norm
concats_list = []
input = Input(shape=input_shape)
origin = input
input = CoordinateChannel2D()(input) if coord_conv == True else input
if encoder_type == "resnet101":
encoder_pretrained = ResNet101(include_top=False,weights='imagenet',input_tensor=input,input_shape=input_shape)
concats_list.append(encoder_pretrained.get_layer('conv1_relu').output) #64, 256*384
concats_list.append(encoder_pretrained.get_layer('conv2_block3_out').output) #256 128*192
concats_list.append(encoder_pretrained.get_layer('conv3_block4_out').output) #512 64*96
concats_list.append(encoder_pretrained.get_layer('conv4_block23_out').output) #1024 32*48
center = encoder_pretrained.layers[-1].output
if encoder_type == "vgg19":
encoder_pretrained = VGG19(include_top=False,weights='imagenet',input_tensor=input,input_shape=input_shape)
concats_list.append(encoder_pretrained.get_layer('block1_conv2').output) #64, 512*768
concats_list.append(encoder_pretrained.get_layer('block2_conv2').output) #128 256*384
concats_list.append(encoder_pretrained.get_layer('block3_conv4').output) #256 128*192
concats_list.append(encoder_pretrained.get_layer('block4_conv4').output) #512 64*96
concats_list.append(encoder_pretrained.get_layer('block5_conv4').output) #512 32*48
center = encoder_pretrained.layers[-1].output
for _ in range(1):
center = Conv2D(1024, (3, 3), padding='same', activation='relu')(center)
center = BatchNormalization()(center)
input = center
# upsampling:
i = 0
if encoder_type == "resnet101":
for f,c in zip([512,256,128,64],concats_list[::-1]):
i+=1
block_name = 'decoder_block' + str(i) + "_conv"
output = self.__decoder_block(input,c,f,block_name)
input = output
output = UpSampling2D((2,2))(output)
output = Conv2D(32, (3, 3), padding='same', activation='relu')(output)
if encoder_type == "vgg19":
for f,c in zip([256,128,64,32,16],concats_list[::-1]):
i+=1
block_name = 'decoder_block' + str(i) + "_conv"
output = self.__decoder_block(input,c,f,block_name)
input = output
final_layer = Conv2D(num_classes, (3, 3), padding='same')(output)
final_layer = Activation("sigmoid")(final_layer)
self.model = Model(inputs=origin, outputs=final_layer)
self.__compile_model()
def freeze_encoder(self,model,encoder_type):
self.model = model
for layer in self.model.layers:
layer.trainable = False
if layer.name == "conv5_block3_out" and encoder_type == "resnet101":
break
if layer.name == "block5_pool" and encoder_type == "vgg19":
break
self.__compile_model()
def unfreeze_encoder(self,model,encoder_type):
self.model = model
for layer in self.model.layers:
layer.trainable = True
if layer.name == "conv5_block3_out" and encoder_type == "resnet101":
break
if layer.name == "block5_pool" and encoder_type == "vgg19":
break
self.__compile_model()
def __decoder_block(self,input,concat_layer,n_feature_maps,block_name):
up = UpSampling2D((2,2))(input)
up = concatenate([concat_layer,up],axis=3)
for i in range(2):
up = Conv2D(n_feature_maps,(3,3),padding='same', activation='relu',name=block_name+str(i+1))(up)
up = BatchNormalization()(up)
return up
def __encoder_block(self,input,n_feature_maps,block_name):
down = input
for i in range(2):
down = Conv2D(n_feature_maps, (3, 3), padding='same', activation='relu',name=block_name+str(i+1))(down)
down = BatchNormalization()(down) if self.batch_norm == True else down
concat_layer = down
down = MaxPooling2D((2, 2), strides=(2, 2))(down)
return down, concat_layer
def __compile_model(self):
self.model.compile(optimizer=Adam(lr=0.0001), loss=bce_dice_loss, metrics=[dice_coeff,iou_score])