-
Notifications
You must be signed in to change notification settings - Fork 0
/
unet.py
210 lines (176 loc) · 9.41 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
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
201
202
203
204
205
206
207
208
209
210
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras import backend as keras_backend
from tensorflow.keras.layers import (Input, Conv2D, Activation, BatchNormalization, Dropout, MaxPooling2D,
UpSampling2D, Cropping2D, concatenate)
from tensorflow.keras.models import load_model
from imgaug.augmentables.segmaps import SegmentationMapsOnImage
class UNet:
def __init__(self, imwidth=128, imheight=128, output_channels=2, autotune=-1):
keras_backend.image_data_format = 'channels_last'
self.model = None
self.output_channels = output_channels
self.imwidth = imwidth
self.imheight = imheight
self.trainset = None
self.valset = None
self.testset = None
self.autotune = autotune
self.seq = None
self.experiment_name = None
def init(self, experiment_name='awesome_project'):
self.experiment_name = experiment_name
@staticmethod
def conv2d_block(input_tensor, filters, kernel_size=(3, 3), padding='valid', batch_norm=True):
x = Conv2D(filters=filters, kernel_size=kernel_size, padding=padding)(input_tensor)
if batch_norm:
x = BatchNormalization()(x)
x = Activation('relu')(x)
# second layer
x = Conv2D(filters=filters, kernel_size=kernel_size,
kernel_initializer='he_normal', padding=padding)(x)
if batch_norm:
x = BatchNormalization()(x)
x = Activation('relu')(x)
return x
def get_unet(self, batch_norm=True, dropout_rate=0.1, optimizer='adam', metrics=None):
if metrics is None:
metrics = ['AUC']
# Input = (N, 512, 512, 1)
x = inputs = Input(shape=(self.imheight, self.imwidth, 1), dtype='float32')
base = self.imheight
# Contracting Path
d1 = self.conv2d_block(input_tensor=x, filters=base / 8, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
p1 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(d1)
p1 = Dropout(rate=dropout_rate)(p1)
d2 = self.conv2d_block(input_tensor=p1, filters=base / 4, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
p2 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(d2)
p2 = Dropout(rate=dropout_rate)(p2)
d3 = self.conv2d_block(input_tensor=p2, filters=base / 2, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
p3 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(d3)
p3 = Dropout(rate=dropout_rate)(p3)
d4 = self.conv2d_block(input_tensor=p3, filters=base, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
p4 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(d4)
p4 = Dropout(rate=dropout_rate)(p4)
# Bottleneck
bn = self.conv2d_block(input_tensor=p4, filters=base * 2, kernel_size=3, padding='same',
batch_norm=batch_norm)
# Expansion Path
u1 = UpSampling2D(size=(2, 2), interpolation='nearest')(bn)
u1 = Conv2D(base, (2, 2), activation='relu', padding='same', kernel_initializer='he_normal')(u1)
cs = (d4.get_shape()[1] - u1.get_shape()[1]) // 2
u1 = concatenate([Cropping2D(cropping=(cs, cs))(d4), u1])
u1 = Dropout(rate=dropout_rate)(u1)
u1 = self.conv2d_block(input_tensor=u1, filters=base, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
u2 = UpSampling2D(size=(2, 2), interpolation='nearest')(u1)
u2 = Conv2D(base / 2, (2, 2), activation='relu', padding='same', kernel_initializer='he_normal')(u2)
cs = (d3.get_shape()[1] - u2.get_shape()[1]) // 2
u2 = concatenate([Cropping2D(cropping=(cs, cs))(d3), u2])
u2 = Dropout(rate=dropout_rate)(u2)
u2 = self.conv2d_block(input_tensor=u2, filters=base / 2, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
u3 = UpSampling2D(size=(2, 2), interpolation='nearest')(u2)
u3 = Conv2D(base / 4, (2, 2), activation='relu', padding='same', kernel_initializer='he_normal')(u3)
cs = (d2.get_shape()[1] - u3.get_shape()[1]) // 2
u3 = concatenate([Cropping2D(cropping=(cs, cs))(d2), u3])
u3 = Dropout(rate=dropout_rate)(u3)
u3 = self.conv2d_block(input_tensor=u3, filters=base / 4, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
u4 = UpSampling2D(size=(2, 2), interpolation='nearest')(u3)
u4 = Conv2D(base / 8, (2, 2), activation='relu', padding='same', kernel_initializer='he_normal')(u4)
cs = (d1.get_shape()[1] - u4.get_shape()[1]) // 2
u4 = concatenate([Cropping2D(cropping=(cs, cs))(d1), u4])
u4 = Dropout(rate=dropout_rate)(u4)
u4 = self.conv2d_block(input_tensor=u4, filters=base / 8, kernel_size=(3, 3), padding='same',
batch_norm=batch_norm)
if self.output_channels > 2:
output_activation = 'softmax'
loss = 'categorical_crossentropy'
elif self.output_channels == 2:
output_activation = 'sigmoid'
loss = 'binary_crossentropy'
else:
raise ValueError
outputs = Conv2D(self.output_channels, (1, 1), activation=output_activation, padding='valid',
kernel_initializer='he_normal')(u4)
self.model = Model(inputs, outputs)
self.model.compile(loss=loss, optimizer=optimizer, metrics=metrics)
def decode_img(self, img):
# convert the compressed string to a 1D uint8 tensor
img = tf.image.decode_png(img, channels=1)
img = tf.image.resize(img, [self.imwidth, self.imheight], method='bilinear')
img = tf.cast(tf.round(img), tf.uint8)
return img
def process_path_double(self, img_path, seg_path):
# load the raw data from the file as a string
img = tf.io.read_file(img_path)
img = self.decode_img(img)
seg = tf.io.read_file(seg_path)
seg = self.decode_img(seg)
# seg = tf.round(seg)
return img, seg
def process_path_single(self, img_path):
img = tf.io.read_file(img_path)
img = self.decode_img(img)
return img
def augment_batch(self, image, segmap):
def augment_image(img, seg):
img = img.numpy().reshape(img.numpy().shape[-3:])
seg = seg.numpy().reshape(seg.numpy().shape[-3:])
shape = img.shape
segg = SegmentationMapsOnImage(seg, shape=shape)
a, b = self.seq.augment(image=img, segmentation_maps=segg)
return a, b.get_arr()
img_shape = tf.shape(image)
segmap_shape = tf.shape(segmap)
image, segmap = tf.py_function(augment_image, [image, segmap], [tf.uint8, tf.uint8])
image = tf.reshape(image, shape=img_shape)
segmap = tf.reshape(segmap, shape=segmap_shape)
return image, segmap
def one_hot_label(self, img, seg):
seg = tf.one_hot(tf.cast(seg, tf.uint8), self.output_channels)
seg = tf.cast(seg, tf.float32)
seg = tf.squeeze(seg, -2)
return img, seg
def normalize_double(self, img, seg):
img = img / 255
seg = seg / 255
return img, seg
def normalize_single(self, img):
return img / 255
def set_seq(self, seq):
self.seq = seq
def process_train(self, im, seg):
self.trainset = tf.data.Dataset.from_tensor_slices((im, seg))
self.trainset = self.trainset.map(self.process_path_double, num_parallel_calls=self.autotune)
self.trainset = self.trainset.map(self.augment_batch, num_parallel_calls=self.autotune)
self.trainset = self.trainset.map(self.normalize_double, num_parallel_calls=self.autotune)
self.trainset = self.trainset.map(self.one_hot_label, num_parallel_calls=self.autotune)
def process_val(self, im, seg):
self.valset = tf.data.Dataset.from_tensor_slices((im, seg))
self.valset = self.valset.map(self.process_path_double, num_parallel_calls=self.autotune)
self.valset = self.valset.map(self.normalize_double, num_parallel_calls=self.autotune)
self.valset = self.valset.map(self.one_hot_label, num_parallel_calls=self.autotune)
def fit(self, trainset, validation_data, steps_per_epoch=100, epochs=1, callbacks=[]):
self.model.fit(trainset, steps_per_epoch=steps_per_epoch, epochs=epochs, callbacks=callbacks,
validation_data=validation_data)
def load_model(self, file_path):
self.model = load_model(file_path)
self.imheight = self.model.input_shape[-3]
self.imwidth = self.model.input_shape[-2]
self.output_channels = self.model.output_shape[-1]
self.autotune = -1
def load_testset(self, im):
if not isinstance(im, list):
im = [im]
self.testset = tf.data.Dataset.from_tensor_slices(im)
self.testset = self.testset.map(self.process_path_single, num_parallel_calls=self.autotune)
self.testset = self.testset.map(self.normalize_single, num_parallel_calls=self.autotune)
# def fit(self, batch, prefetch, repeat, epochs, callbacks):
# self.model.fit(self.trainset.batch(batch).prefetch(prefetch).repeat(repeat),
# validation_data=self.valset.batch(batch), epochs=epochs, callbacks=callbacks)