-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathunet1.py
159 lines (119 loc) · 8.13 KB
/
unet1.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
from __future__ import print_function
import os
import numpy as np
from keras.models import Model
from keras.layers import Input, concatenate, Conv1D, MaxPooling1D, Conv2DTranspose,Lambda,BatchNormalization,LSTM
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint
from keras import backend as K
import keras
import tensorflow as tf
from keras.layers import ZeroPadding1D
K.set_image_data_format('channels_last') # TF dimension ordering in this code
size= 4096*1024
channel=5
batch_size=32
ss=10
def crossentropy_cut(y_true,y_pred):
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
y_pred_f= tf.clip_by_value(y_pred_f, 1e-7, (1. - 1e-7))
# mask=K.cast(K.greater_equal(y_true_f,-0.5),dtype='float32')
mask=K.greater_equal(y_true_f,-0.5)
# out = -(y_true_f * K.log(y_pred_f)*mask + (1.0 - y_true_f) * K.log(1.0 - y_pred_f)*mask)
# out=K.mean(out)
losses = -(y_true_f * K.log(y_pred_f) + (1.0 - y_true_f) * K.log(1.0 - y_pred_f))
losses = tf.boolean_mask(losses, mask)
masked_loss = tf.reduce_mean(losses)
return masked_loss
def dice_coef(y_true, y_pred):
y_true_f = K.flatten(y_true)
def Conv1DTranspose(input_tensor, filters, kernel_size, strides=2, padding='same'):
x = Lambda(lambda x: K.expand_dims(x, axis=2))(input_tensor)
x = Conv2DTranspose(filters=filters, kernel_size=(kernel_size, 1), strides=(strides, 1), padding=padding)(x)
x = Lambda(lambda x: K.squeeze(x, axis=2))(x)
return x
def dice_coef(y_true, y_pred):
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
mask=K.cast(K.greater_equal(y_true_f,-0.5),dtype='float32')
intersection = K.sum(y_true_f * y_pred_f * mask)
return (2. * intersection + ss) / (K.sum(y_true_f * mask) + K.sum(y_pred_f * mask) + ss)
def dice_coef_loss(y_true, y_pred):
return -dice_coef(y_true, y_pred)
def get_unet():
inputs = Input((size, channel)) #4096*1024
print(inputs.shape)
conv01 = BatchNormalization()(Conv1D(15, 7, activation='relu', padding='same')(inputs))
conv01 = BatchNormalization()(Conv1D(15, 7, activation='relu', padding='same')(conv01))
pool01 = MaxPooling1D(pool_size=2)(conv01) #4096*512
conv0 = BatchNormalization()(Conv1D(18, 7, activation='relu', padding='same')(pool01))#+8
conv0 = BatchNormalization()(Conv1D(18, 7, activation='relu', padding='same')(conv0))
pool0 = MaxPooling1D(pool_size=4)(conv0) #4096*128
conv1 = BatchNormalization()(Conv1D(21, 7, activation='relu', padding='same')(pool0))#+8
conv1 = BatchNormalization()(Conv1D(21, 7, activation='relu', padding='same')(conv1))
pool1 = MaxPooling1D(pool_size=4)(conv1) #4096*32
# conv2 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(pool1))#+16
# conv2 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(conv2))
# pool2 = MaxPooling1D(pool_size=2)(conv2) #4096*8
conv3 = BatchNormalization()(Conv1D(30, 7, activation='relu', padding='same')(pool1))#+16
conv3 = BatchNormalization()(Conv1D(30, 7, activation='relu', padding='same')(conv3))
pool3 = MaxPooling1D(pool_size=4)(conv3) #4096*8
# conv4 = BatchNormalization()(Conv1D(112, 7, activation='relu', padding='same')(pool3))#+32
# conv4 = BatchNormalization()(Conv1D(112, 7, activation='relu', padding='same')(conv4))
# pool4 = MaxPooling1D(pool_size=2)(conv4) #4096*2
conv5 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(pool3))#+32
conv5 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(conv5))
pool5 = MaxPooling1D(pool_size=4)(conv5) #4096*2
# conv6 = BatchNormalization()(Conv1D(208, 7, activation='relu', padding='same')(pool5))#+64
# conv6 = BatchNormalization()(Conv1D(208, 7, activation='relu', padding='same')(conv6))
# pool6 = MaxPooling1D(pool_size=2)(conv6) #2048
conv7 = BatchNormalization()(Conv1D(120, 7, activation='relu', padding='same')(pool5))#+64
conv7 = BatchNormalization()(Conv1D(120, 7, activation='relu', padding='same')(conv7))
pool7 = MaxPooling1D(pool_size=4)(conv7) #2048
# conv8 = BatchNormalization()(Conv1D(400, 7, activation='relu', padding='same')(pool7))#+128
# conv8 = BatchNormalization()(Conv1D(400, 7, activation='relu', padding='same')(conv8))
# pool8 = MaxPooling1D(pool_size=2)(conv8) #512
conv9 = BatchNormalization()(Conv1D(240, 7, activation='relu', padding='same')(pool7))#+128
conv9 = BatchNormalization()(Conv1D(240, 7, activation='relu', padding='same')(conv9))
pool9 = MaxPooling1D(pool_size=4)(conv9) #512
conv10 = BatchNormalization()(Conv1D(480, 7, activation='relu', padding='same')(pool9))#+496
conv10 = BatchNormalization()(Conv1D(480, 7, activation='relu', padding='same')(conv10))
#lstm0 = CuDNNLSTM(1900,return_sequences=True)(conv10)
up11 = concatenate([Conv1DTranspose(conv10,240, 4, strides=4, padding='same'), conv9], axis=2)
conv11 = BatchNormalization()(Conv1D(240, 7, activation='relu', padding='same')(up11))
conv11 = BatchNormalization()(Conv1D(240, 7, activation='relu', padding='same')(conv11)) #1024
# up12 = concatenate([Conv1DTranspose(conv11,400, 2, strides=2, padding='same'), conv8], axis=2)
# conv12 = BatchNormalization()(Conv1D(400, 7, activation='relu', padding='same')(up12))
# conv12 = BatchNormalization()(Conv1D(400, 7, activation='relu', padding='same')(conv12)) #1024
up13 = concatenate([Conv1DTranspose(conv11,120, 4, strides=4, padding='same'), conv7], axis=2)
conv13 = BatchNormalization()(Conv1D(120, 7, activation='relu', padding='same')(up13))
conv13 = BatchNormalization()(Conv1D(120, 7, activation='relu', padding='same')(conv13)) #4096
# up14 = concatenate([Conv1DTranspose(conv13,208, 2, strides=2, padding='same'), conv6], axis=2)
# conv14 = BatchNormalization()(Conv1D(208, 7, activation='relu', padding='same')(up14))
# conv14 = BatchNormalization()(Conv1D(208, 7, activation='relu', padding='same')(conv14)) #4096
up15 = concatenate([Conv1DTranspose(conv13,60, 4, strides=4, padding='same'), conv5], axis=2)
conv15 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(up15))
conv15 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(conv15)) #4096*4
# up16 = concatenate([Conv1DTranspose(conv15,112, 2, strides=2, padding='same'), conv4], axis=2)
# conv16 = BatchNormalization()(Conv1D(112, 7, activation='relu', padding='same')(up16))
# conv16 = BatchNormalization()(Conv1D(112, 7, activation='relu', padding='same')(conv16)) #4096*4
up17 = concatenate([Conv1DTranspose(conv15,30, 4, strides=4, padding='same'), conv3], axis=2)
conv17 = BatchNormalization()(Conv1D(30, 7, activation='relu', padding='same')(up17))
conv17 = BatchNormalization()(Conv1D(30, 7, activation='relu', padding='same')(conv17)) #4096*16
# up18 = concatenate([Conv1DTranspose(conv17,60, 2, strides=2, padding='same'), conv2], axis=2)
# conv18 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(up18))
# conv18 = BatchNormalization()(Conv1D(60, 7, activation='relu', padding='same')(conv18)) #4096*16
up19 = concatenate([Conv1DTranspose(conv17,21, 4, strides=4, padding='same'), conv1], axis=2)
conv19 = BatchNormalization()(Conv1D(21, 7, activation='relu', padding='same')(up19))
conv19 = BatchNormalization()(Conv1D(21, 7, activation='relu', padding='same')(conv19)) #4096*64
up20 = concatenate([Conv1DTranspose(conv19,18, 4, strides=4, padding='same'), conv0], axis=2)
conv20 = BatchNormalization()(Conv1D(18, 7, activation='relu', padding='same')(up20))
conv20 = BatchNormalization()(Conv1D(18, 7, activation='relu', padding='same')(conv20)) #4096*64
up21 = concatenate([Conv1DTranspose(conv20,15, 2, strides=2, padding='same'), conv01], axis=2)
conv21 = BatchNormalization()(Conv1D(15, 7, activation='relu', padding='same')(up21))
conv21 = BatchNormalization()(Conv1D(15, 7, activation='relu', padding='same')(conv21)) #4096*256
conv22 = Conv1D(1, 1, activation='sigmoid')(conv21)
model = Model(inputs=[inputs], outputs=[conv22])
model.compile(optimizer=Adam(lr=1e-4,beta_1=0.9, beta_2=0.999,decay=1e-5), loss=crossentropy_cut, metrics=[dice_coef])
return model