-
Notifications
You must be signed in to change notification settings - Fork 0
/
unet_model.py
479 lines (401 loc) · 18 KB
/
unet_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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
""" Trains U-Net
"""
import numpy as np
from keras.models import Model
from keras.models import load_model
from keras.layers import (Input,
merge,
Convolution2D,
MaxPooling2D,
Deconvolution2D,
UpSampling2D,
Dropout,
Cropping2D,
Lambda,
Activation,
merge)
from keras.layers.advanced_activations import PReLU
from keras.layers.pooling import GlobalAveragePooling2D
from keras.layers.normalization import BatchNormalization
from keras.optimizers import Adam
from keras.callbacks import (ModelCheckpoint,
LearningRateScheduler)
from keras import backend as K
from custom_image import (ImageDataGenerator,
standardize,
random_transform)
from custom_layers import (CroppingChannels,
DepthSoftmax)
from net_utils import *
def downblock_seg(input_net, nb_filters, init):
db = BatchNormalization()(input_net)
db = PReLU()(db)
db = Convolution2D(nb_filters, 3, 3, border_mode='same', init=init, subsample=(2,2))(db)
db_2 = BatchNormalization()(db)
db_2 = PReLU()(db_2)
db_2 = Convolution2D(nb_filters, 3, 3, border_mode='same', init=init)(db_2)
db_2 = merge([db, db_2], mode='sum')
return db_2
def upblock_seg(input_net, nb_filters, deconv_output, init='he_normal'):
sb = Convolution2D(nb_filters, 1, 1, init=init)(input_net)
sb = BatchNormalization()(sb)
sb = PReLU()(sb)
sb = Deconvolution2D(nb_filters, 2, 2,
output_shape=deconv_output,
subsample=(2, 2),
init=init)(sb)
sb = BatchNormalization()(sb)
sb = PReLU()(sb)
return sb
def merge_seg(input_net, skip_net, nb_filters, init='he_norma'):
ms = merge([input_net, skip_net], mode='concat', concat_axis=3)
ms = Convolution2D(nb_filters, 3, 3, border_mode='same', init=init)(ms)
ms = BatchNormalization()(ms)
ms = PReLU()(ms)
return ms
def downblock_vgg(input_net, nb_filters, init):
db = Convolution2D(nb_filters, 3, 3, activation='relu', border_mode='valid', init=init)(input_net)
db = BatchNormalization()(db)
db = Convolution2D(nb_filters, 3, 3, activation='relu', border_mode='valid', init=init)(db)
db = BatchNormalization()(db)
db_pooled = MaxPooling2D(pool_size=(2, 2))(db)
return (db, db_pooled)
def upblock_vgg(input_net,
skip_net=None,
deconv_output=None,
crop_width=None,
nb_filters=64,
init='he_normal',
name=None):
"""Provides a vgg-like up-block of model steps
Args:
input_net: previous model output which is fed in
skip_net: previous model output which is skips across and is appended to filters,
if None then ignored
deconv_output: output of the transposed convolution layer
crop_width: amount to crop skip_net by to match with deconv_output
nb_filters: number of filters
init: weight initialization
Output:
ub: complete model which describes up-block
"""
if name:
name_crop2D = name + '_crop2D'
name_deconv = name + '_deconv2D'
name_bn_1 = name + '_bn_1'
name_conv2D_1 = name + '_conv2D_1'
name_bn_2 = name + '_bn_2'
name_conv2D_2 = name + '_conv2d_2'
name_bn_3 = name + '_bn_3'
if skip_net is not None:
ub = Cropping2D(cropping=((crop_width, crop_width), (crop_width, crop_width)),
name=name_crop2D)(skip_net)
deconv = Deconvolution2D(nb_filters, 2, 2,
output_shape=deconv_output,
subsample=(2, 2),
init=init,
name=name_deconv)(input_net)
ub = merge([deconv, ub], mode='concat', concat_axis=3)
else:
ub = Deconvolution2D(nb_filters, 2, 2,
output_shape=deconv_output,
subsample=(2, 2),
init=init,
name=name_deconv)(input_net)
ub = BatchNormalization(name=name_bn_1)(ub)
ub = Convolution2D(nb_filters, 3, 3,
activation='relu',
border_mode='valid',
init=init,
name=name_conv2D_1)(ub)
ub = BatchNormalization(name=name_bn_2)(ub)
ub = Convolution2D(nb_filters, 3, 3,
activation='relu',
border_mode='valid',
init=init,
name=name_conv2D_2)(ub)
ub = BatchNormalization(name=name_bn_3)(ub)
return ub
def get_aenet(batch_size=1,
rows=572,
cols=572,
num_channels=20,
nb_filters=64,
loss='jaccard',
net_type='vgg',
init='he_normal',
lr=1e-3):
"""Returns keras model of a VGG-like autoencoder for a first pass.
Args:
batch_size: batch_size of samples
rows: rows of image
cols: cols of image
num_channels: number channels in input images
lr: learning rate of optimizer
nb_filters: base number of filters used (gets increased in successive
layers)
loss: One of 'jaccard', 'logjaccard', or 'logloss', gives choice of
objective fn
init: weight initialization, default 'he_normal'
Output:
model: CNN model
"""
# Tensorflow does not allow using None for batch_size in deconv layers
# we can input it symbolically or we can just hard-code batch size
# see https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/vf8eH9YMwVA
# and https://github.com/fchollet/keras/issues/3478
# Tensorflow does not allow using None for batch_size in deconv layers
# we can input it symbolically or we can just hard-code batch size
# see https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/vf8eH9YMwVA
# and https://github.com/fchollet/keras/issues/3478
inputs = Input((rows, cols, num_channels))
# keeps only the 4 high-res channels (RGB and P image)
hires_channels = CroppingChannels(cropping=(0,17))(inputs)
(conv1_skip, conv1) = downblock_vgg(hires_channels, nb_filters, init)
(conv2_skip, conv2) = downblock_vgg(conv1, nb_filters*2, init)
(conv3_skip, conv3) = downblock_vgg(conv2, nb_filters*4, init)
(conv4_skip, conv4) = downblock_vgg(conv3, nb_filters*8, init)
(conv5_skip, conv5) = downblock_vgg(conv4, nb_filters*16, init)
conv6 = upblock_vgg(conv5_skip,
deconv_output=[batch_size, 56, 56, nb_filters*8],
crop_width=4,
nb_filters=nb_filters*8,
init=init)
conv7 = upblock_vgg(conv6,
deconv_output=[batch_size, 104, 104, nb_filters*4],
crop_width=16,
nb_filters=nb_filters*4,
init=init)
conv8 = upblock_vgg(conv7,
deconv_output=[batch_size, 200, 200, nb_filters*2],
crop_width=40,
nb_filters=nb_filters*2,
init=init)
conv9 = upblock_vgg(conv8,
deconv_output=[batch_size, 392, 392, nb_filters],
crop_width=88,
nb_filters=nb_filters,
init=init)
# this is explicity an auto-encoder, so we make sure that they are the same
conv10 = Convolution2D(3, 1, 1, activation='sigmoid')(conv9)
model = Model(input=inputs, output=conv10)
model.compile(optimizer=Adam(lr=lr), loss='binary_crossentropy')
return model
def get_unet(batch_size=1,
shape_in = (572, 572, 20),
shape_out = (388, 388, 11),
lr=1e-3,
nb_filters=64,
loss='jaccard',
net_type='vgg',
init='he_normal',
ae_weights=None,
classifier_weights=None,
mask_weight=None,
end_activation='softmax'):
"""Returns keras model of a small U-Net CNN for prototyping. See:
https://arxiv.org/pdf/1505.04597v1.pdf
Args:
batch_size: batch_size of samples
rows: rows of image
cols: cols of image
num_channels: number channels in input images
lr: learning rate of optimizer
nb_filters: base number of filters used (gets increased in successive
layers)
loss: One of 'jaccard', 'logjaccard', or 'logloss', gives choice of
objective fn
init: weight initialization, default 'he_normal'
ae_weights: filepath to pre-trained weights of autoencoder
classifier_weights: filepath to pretrained weights of classifier
(the 'encoding' portion of the net)
Output:
model: CNN model
"""
# Tensorflow does not allow using None for batch_size in deconv layers
# we can input it symbolically or we can just hard-code batch size
# see https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/vf8eH9YMwVA
# and https://github.com/fchollet/keras/issues/3478
# Tensorflow does not allow using None for batch_size in deconv layers
# we can input it symbolically or we can just hard-code batch size
# see https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/vf8eH9YMwVA
# and https://github.com/fchollet/keras/issues/3478
# make model from scratch if not passing in weights
if classifier_weights is None:
inputs = Input(shape_in)
# keeps only the 3 high-res channels (RGB and P image)
hires_channels = CroppingChannels((0,17))(inputs)
(conv1_skip, conv1) = downblock_vgg(inputs, nb_filters, init)
(conv2_skip, conv2) = downblock_vgg(conv1, nb_filters*2, init)
(conv3_skip, conv3) = downblock_vgg(conv2, nb_filters*4, init)
(conv4_skip, conv4) = downblock_vgg(conv3, nb_filters*8, init)
(conv5_skip, conv5) = downblock_vgg(conv4, nb_filters*16, init)
# If we pass in weights, we build out model around the pretrained weights
else:
vgg = load_model(classifier_weights,
custom_objects={'CroppingChannels': CroppingChannels})
inputs = vgg.get_layer('input_1').output
conv1_skip = vgg.get_layer('batchnormalization_2').output
conv2_skip = vgg.get_layer('batchnormalization_4').output
conv3_skip = vgg.get_layer('batchnormalization_6').output
conv4_skip = vgg.get_layer('batchnormalization_8').output
conv5_skip = vgg.get_layer('batchnormalization_10').output
conv1 = vgg.get_layer('maxpooling2d_1').output
conv2 = vgg.get_layer('maxpooling2d_2').output
conv3 = vgg.get_layer('maxpooling2d_3').output
conv4 = vgg.get_layer('maxpooling2d_4').output
conv6 = upblock_vgg(conv5_skip, conv4_skip,
deconv_output=[batch_size, 56, 56, nb_filters*8],
crop_width=4,
nb_filters=nb_filters*8,
init=init,
name='ub_vgg_1')
conv7 = upblock_vgg(conv6, conv3_skip,
deconv_output=[batch_size, 104, 104, nb_filters*4],
crop_width=16,
nb_filters=nb_filters*4,
init=init,
name='ub_vgg_2')
conv8 = upblock_vgg(conv7, conv2_skip,
deconv_output=[batch_size, 200, 200, nb_filters*2],
crop_width=40,
nb_filters=nb_filters*2,
init=init,
name='ub_vgg_3')
conv9 = upblock_vgg(conv8, conv1_skip,
deconv_output=[batch_size, 392, 392, nb_filters],
crop_width=88,
nb_filters=nb_filters,
init=init,
name='ub_vgg_4')
# adds in the other 16 channels, and crops them appropriately
# Done like this, we are essentially using the RGB channels for feature
# finding, and then adding in quasi-pixel-based classification for the
# final segmentation map
lowres_channels = Cropping2D(cropping=((92,92), (92,92)),
name='lowres_spatialcrop')(inputs)
lowres_channels = CroppingChannels(cropping=(3,0),
name='lowres_channelcrop')(lowres_channels)
# appends them to current stack
conv10 = merge([conv9, lowres_channels], mode='concat', concat_axis=3)
conv10 = Convolution2D(nb_filters*2, 1, 1,
name='conv10_conv2d_1')(conv10)
conv10 = Convolution2D(nb_filters*2, 1, 1,
name='conv10_conv2d_2')(conv10)
conv11 = Convolution2D(shape_out[2], 1, 1,
name='conv11_conv2d_1')(conv10)
if end_activation == 'softmax':
conv11 = DepthSoftmax()(conv11)
elif end_activation == 'sigmoid':
conv11 = Activation('sigmoid')(conv11)
model = Model(input=inputs, output=conv11)
if mask_weight is not None:
if loss == 'jaccard':
loss_fn = jaccard_coef_loss
elif loss == 'logloss':
loss_fn = pixelwise_logloss
elif loss == 'logjaccard':
loss_fn = jaccard_coef_logloss
else:
raise ValueError('The loss must be one of \'jaccard\','
+'\'logjaccard\', or \'logloss\'')
else:
# if the user passes a list of mask weights, we use a modified
# loss function
loss_fn = partial(jaccard_coef_loss_weighted, weights=mask_weight)
model.compile(optimizer=Adam(lr=lr),
loss=loss_fn,
metrics=[jaccard_coef])
return model
def get_classnet(batch_size=8,
shape_in=(572, 572, 20),
lr=1e-3,
nb_filters=64,
init='he_normal'):
"""Returns keras model of a small U-Net CNN for prototyping. See:
https://arxiv.org/pdf/1505.04597v1.pdf
Args:
Output:
model: CNN model
"""
# Tensorflow does not allow using None for batch_size in deconv layers
# we can input it symbolically or we can just hard-code batch size
# see https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/vf8eH9YMwVA
# and https://github.com/fchollet/keras/issues/3478
inputs = Input(shape_in)
# keeps only the 4 high-res channels (RGB and P image)
hires_channels = CroppingChannels(cropping=(0,17))(inputs)
(conv1_skip, conv1) = downblock_vgg(hires_channels, nb_filters, init)
(conv2_skip, conv2) = downblock_vgg(conv1, nb_filters*2, init)
(conv3_skip, conv3) = downblock_vgg(conv2, nb_filters*4, init,)
(conv4_skip, conv4) = downblock_vgg(conv3, nb_filters*8, init)
(conv5_skip, conv5) = downblock_vgg(conv4, nb_filters*16, init)
conv6 = Convolution2D(nb_filters*16, 1, 1, activation='relu', init=init)(conv5_skip)
conv6 = Convolution2D(11, 1, 1, activation='relu', init=init)(conv6)
conv6 = GlobalAveragePooling2D()(conv6)
conv6 = Activation('softmax')(conv6)
model = Model(input=inputs, output=conv6)
model.compile(optimizer=Adam(lr=lr), loss='categorical_crossentropy', metrics=['categorical_accuracy'])
return model
def get_segnet(batch_size=1,
shape_in = (512, 512, 20),
shape_out = (512, 512, 11),
lr=5e-5,
nb_filters=64,
loss='jaccard',
init='he_normal',
end_activation='softmax'):
"""Returns keras model of a sophisticated u-net type for medical
segmentation, see:
https://arxiv.org/pdf/1701.03056.pdf
Args:
batch_size: batch_size of samples
shape_in: input shape
shape_out: output shape
num_channels: number channels in input images
lr: learning rate of optimizer
nb_filters: base number of filters used (gets increased in successive
layers)
loss: One of 'jaccard', 'logjaccard', or 'logloss', gives choice of
objective fn
init: weight initialization, default 'he_normal'
Output:
model: CNN model
"""
inputs = Input(shape_in)
# beginning convolutions
conv_1 = Convolution2D(nb_filters, 3, 3, border_mode='same', init=init)(inputs)
conv_2 = Convolution2D(nb_filters, 3, 3, border_mode='same', init=init)(conv_1)
# down-sampling
db_1 = downblock_seg(conv_2, nb_filters, init)
db_2 = downblock_seg(db_1, nb_filters*2, init)
db_3 = downblock_seg(db_2, nb_filters*4, init)
# up-sampling
deconv_output = [batch_size, shape_in[0]//4, shape_in[1]//4, nb_filters*4]
ub_1 = upblock_seg(db_3, nb_filters*4, deconv_output=deconv_output, init=init)
ms_1 = merge_seg(ub_1, db_2, nb_filters*2, init)
deconv_output = [batch_size, shape_in[0]//2, shape_in[1]//2, nb_filters*2]
ub_2 = upblock_seg(ms_1, nb_filters*2, deconv_output=deconv_output, init=init)
ms_2 = merge_seg(ub_2, db_1, nb_filters*2, init)
deconv_output = [batch_size, shape_in[0], shape_in[1], nb_filters]
ub_3 = upblock_seg(ms_2, nb_filters, deconv_output=deconv_output, init=init)
ms_3 = merge_seg(ub_3, conv_2, nb_filters, init)
# bringing outputs of different layers together in map
final_1 = Convolution2D(shape_out[2], 1, 1, init=init)(ms_1)
final_1 = UpSampling2D(size=(2,2))(final_1)
final_2 = Convolution2D(shape_out[2], 1, 1, init=init)(ms_2)
final_2 = merge([final_2, final_1], mode='sum')
final_2 = UpSampling2D(size=(2,2))(final_2)
final_3 = Convolution2D(shape_out[2], 1, 1, init=init)(ms_3)
final_3 = merge([final_3, final_2], mode='sum')
#activated = Activation('softmax')(final_3)
if end_activation='softmax':
activated = DepthSoftmax()(final_3)
else:
activated = Activation('sigmoid')(final_3)
model = Model(input=inputs, output=activated)
loss_fn = jaccard_coef_loss
model.compile(optimizer=Adam(lr=lr),
loss=loss_fn,
metrics=[jaccard_coef])
return model