-
Notifications
You must be signed in to change notification settings - Fork 12
/
model.py
403 lines (315 loc) · 16.6 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
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
import numpy as np
import matplotlib
matplotlib.use('Agg')
import sys
sys.setrecursionlimit(10000)
from keras.optimizers import Adam
from keras.models import Model
from keras.layers import Input, merge, Lambda, LeakyReLU, MaxPooling2D
from keras.layers.convolutional import Conv2D, UpSampling2D
from keras import backend as K
from keras.layers.core import Dense, Activation, Flatten
from SpatialTransformerLayer import SpatialTransformer
class Multimodel(object):
'''
Class for constructing a neural network model as described in
T. Joyce, A. Chartsias, S.A. Tsaftaris, 'Robust Multi-Modal MR Image Synthesis,' MICCAI 2017
The compiled Keras model inputs and outputs are the following:
inputs: list of numpy data arrays, one for each modality
outputs: list containing numpy arrays for each output modality, 2 zero numpy arrays (one used for variance minimisation,
the other as a dummy value since the last output of the model contains latent representations)
The model is 2D, so the input numpy arrays are of size (<num_images>, <channels>, <height>, <width>)
Example usage:
m = Multimodel(['T1','T2'], ['DWI', 'VFlair'], {'DWI': 1.0, 'VFlair': 1.0, 'concat': 1.0}, 16, 1, True, 'max', True, True)
m.build()
'''
def __init__(self, input_modalities, output_modalities, output_weights, latent_dim, channels, spatial_transformer,
common_merge, ind_outs, fuse_outs):
self.input_modalities = input_modalities
self.output_modalities = output_modalities
self.latent_dim = latent_dim
self.channels = channels
self.common_merge = common_merge
self.output_weights = output_weights
self.spatial_transformer = spatial_transformer
self.ind_outs = ind_outs
self.fuse_outs = fuse_outs
self.num_emb = len(input_modalities) + 1
if spatial_transformer:
self.H, self.W = 112, 80 # Width/Height for ISLES2015 dataset
else:
self.H, self.W = None, None
def encoder_maker(self, modality):
inp = Input(shape=(self.channels, self.H, self.W), name='enc_' + modality + '_input')
conv = Conv2D(32, 3, padding='same', name='enc_' + modality + '_conv1')(inp)
act = LeakyReLU()(conv)
conv = Conv2D(32, 3, padding='same', name='enc_' + modality + '_conv2')(act)
act1 = LeakyReLU()(conv)
# downsample 1st level
pool = MaxPooling2D(pool_size=(2, 2))(act1)
conv = Conv2D(64, 3, padding='same', name='enc_' + modality + '_conv3')(pool)
act = LeakyReLU()(conv)
conv = Conv2D(64, 3, padding='same', name='enc_' + modality + '_conv4')(act)
act2 = LeakyReLU()(conv)
# downsample 2nd level
pool = MaxPooling2D(pool_size=(2, 2))(act2)
conv = Conv2D(128, 3, padding='same', name='enc_' + modality + '_conv5')(pool)
act = LeakyReLU()(conv)
conv = Conv2D(128, 3, padding='same', name='enc_' + modality + '_conv6')(act)
act = LeakyReLU()(conv)
# upsample 2nd level
ups = UpSampling2D(size=(2, 2))(act)
conv = Conv2D(64, 3, padding='same', name='enc_' + modality + '_conv7')(ups)
skip = merge([act2, conv], mode='concat', concat_axis=1, name='enc_' + modality + '_skip1')
conv = Conv2D(64, 3, padding='same', name='enc_' + modality + '_conv8')(skip)
act = LeakyReLU()(conv)
conv = Conv2D(64, 3, padding='same', name='enc_' + modality + '_conv9')(act)
act = LeakyReLU()(conv)
# upsample 2nd level
ups = UpSampling2D(size=(2, 2))(act)
conv = Conv2D(32, 3, padding='same', name='enc_' + modality + '_conv10')(ups)
skip = merge([act1, conv], mode='concat', concat_axis=1, name='enc_' + modality + '_skip2')
conv = Conv2D(32, 3, padding='same', name='enc_' + modality + '_conv11')(skip)
act = LeakyReLU()(conv)
conv = Conv2D(32, 3, padding='same', name='enc_' + modality + '_conv12')(act)
act = LeakyReLU()(conv)
conv_ld = self.latent_dim / 2 if self.common_merge == 'hemis' else self.latent_dim
conv = Conv2D(conv_ld, 3, padding='same', name='enc_' + modality + '_conv13')(act)
lr = LeakyReLU()(conv)
return inp, lr
def decoder_maker(self, modality):
inp = Input(shape=(self.latent_dim, None, None), name='dec_' + modality + '_input')
conv = Conv2D(32, 3, padding='same', activation='relu', name='dec_' + modality + '_conv1')(inp)
conv = Conv2D(32, 3, padding='same', activation='relu', name='dec_' + modality + '_conv2')(conv)
skip = merge([inp, conv], mode='concat', concat_axis=1, name='dec_' + modality + '_skip1')
conv = Conv2D(32, 3, padding='same', activation='relu', name='dec_' + modality + '_conv3')(skip)
conv = Conv2D(32, 3, padding='same', activation='relu', name='dec_' + modality + '_conv4')(conv)
skip = merge([skip, conv], mode='concat', concat_axis=1, name='dec_' + modality + '_skip2')
conv = Conv2D(1, 1, padding='same', activation='relu', name='dec_' + modality + '_conv5')(skip)
model = Model(input=inp, output=conv, name='decoder_' + modality)
return model
def get_embedding_distance_outputs(self, embeddings):
if len(self.inputs) == 1:
print 'Skipping embedding distance outputs for unimodal model'
return []
outputs = list()
ind_emb = embeddings[:-1]
weighted_rep = embeddings[-1]
all_emb_flattened = [new_flatten(emb) for emb in ind_emb]
concat_emb = merge(all_emb_flattened, mode='concat', concat_axis=1, name='em_concat')
concat_emb.name = 'em_concat'
outputs.append(concat_emb)
print 'making output: em_concat', concat_emb.type, concat_emb.name
fused_emb = new_flatten(weighted_rep, name='em_fused')
fused_emb.name = 'em_fused'
outputs.append(fused_emb)
return outputs
# HeMIS based fusion:
# M. Havaei, N. Guizard, N. Chapados, and Y. Bengio, “HeMIS: Hetero- modal image segmentation,”
# in MICCAI. Springer, 2016, pp. 469–477
def hemis(self, ind_emb):
if len(self.input_modalities) == 1:
combined_emb1 = ind_emb[0]
combined_emb2 = K.zeros_like(ind_emb[0]) # if we only have one input the variance is 0
else:
combined_emb1 = merge(ind_emb, mode='ave', name='combined_em_ave',
output_shape=(self.latent_dim / 2, None, None))
combined_emb2 = merge(ind_emb, mode=var, name='combined_em_var',
output_shape=(self.latent_dim / 2, None, None))
combined_emb = merge([combined_emb1, combined_emb2], mode='concat', concat_axis=1, name='combined_em',
output_shape=(self.latent_dim, None, None))
new_ind_emb = []
for i, emb in enumerate(ind_emb):
new_ind_emb.append(merge([emb, zeros_for_var(emb)], mode='concat', concat_axis=1, name='emb_' + str(i),
output_shape=(self.latent_dim, None, None)))
ind_emb = new_ind_emb
all_emb = ind_emb + [combined_emb]
return all_emb
def build(self):
print 'Latent dimensions: ' + str(self.latent_dim)
encoders = [self.encoder_maker(m) for m in self.input_modalities]
ind_emb = [lr for (input, lr) in encoders]
self.org_ind_emb = [lr for (input, lr) in encoders]
self.inputs = [input for (input, lr) in encoders]
# apply spatial transformer
if self.spatial_transformer:
print 'Adding a spatial transformer layer'
input_shape = (self.latent_dim, self.H, self.W)
tpn = tpn_maker(input_shape)
mod1 = ind_emb[0]
aligned_ind_emb = [mod1]
for mod in ind_emb[1:]:
aligned_mod = merge([tpn([mod1, mod]), mod], mode=STMerge, output_shape=input_shape)
aligned_ind_emb.append(aligned_mod)
ind_emb = aligned_ind_emb
if self.common_merge == 'hemis':
self.all_emb = self.hemis(ind_emb)
else:
assert self.common_merge == 'max' or self.common_merge == 'ave' or self.common_merge == 'rev_loss'
print 'Fuse latent representations using ' + str(self.common_merge)
cm = 'max' if self.common_merge == 'rev_loss' else self.common_merge
weighted_rep = merge(ind_emb, mode=cm, name='combined_em') if len(self.inputs) > 1 else ind_emb[0]
self.all_emb = ind_emb + [weighted_rep]
self.decoders = [self.decoder_maker(m) for m in self.output_modalities]
outputs = get_decoder_outputs(self.output_modalities, self.decoders, self.all_emb)
# this is for minimizing the distance between the individual embeddings
outputs += self.get_embedding_distance_outputs(self.all_emb)
print 'all outputs: ', [o.name for o in outputs]
out_dict = {'em_%d_dec_%s' % (emi, dec): mae for emi in range(self.num_emb) for dec in self.output_modalities}
get_indiv_weight = lambda mod: self.output_weights[mod] if self.ind_outs else 0.0
get_fused_weight = lambda mod: self.output_weights[mod] if self.fuse_outs else 0.0
loss_weights = {}
for dec in self.output_modalities:
for emi in range(self.num_emb - 1):
loss_weights['em_%d_dec_%s' % (emi, dec)] = get_indiv_weight(dec)
loss_weights['em_%d_dec_%s' % (self.num_emb - 1, dec)] = get_fused_weight(dec)
if len(self.inputs) > 1:
if self.common_merge == 'rev_loss':
out_dict['em_concat'] = mae
else:
out_dict['em_concat'] = embedding_distance
loss_weights['em_concat'] = self.output_weights['concat']
out_dict['em_fused'] = embedding_distance
loss_weights['em_fused'] = 0.0
print 'output dict: ', out_dict
print 'loss weights: ', loss_weights
self.model = Model(input=self.inputs, output=outputs)
self.model.compile(optimizer=Adam(lr=0.0001), loss=out_dict, loss_weights=loss_weights)
def get_inputs(self, modalities):
return [self.inputs[self.input_modalities.index(mod)] for mod in modalities]
def get_embeddings(self, modalities):
assert set(modalities).issubset(set(self.input_modalities))
ind_emb = [self.all_emb[self.input_modalities.index(mod)] for mod in modalities]
org_ind_emb = [self.org_ind_emb[self.input_modalities.index(mod)] for mod in modalities]
if self.common_merge == 'hemis':
combined_emb1 = merge(org_ind_emb, mode='ave', name='combined_em_ave',
output_shape=(self.latent_dim / 2, None, None))
combined_emb2 = merge(org_ind_emb, mode=var, name='combined_em_var',
output_shape=(self.latent_dim / 2, None, None))
combined_emb = merge([combined_emb1, combined_emb2], mode='concat', concat_axis=1, name='combined_em',
output_shape=(self.latent_dim, None, None))
new_ind_emb = []
for i, emb in enumerate(org_ind_emb):
new_ind_emb.append(merge([emb, zeros_for_var(emb)], mode='concat', concat_axis=1, name='pemb_' + str(i),
output_shape=(self.latent_dim, None, None)))
ind_emb = new_ind_emb
return ind_emb + [combined_emb]
else:
if len(ind_emb) > 1:
fused_emb = merge(ind_emb, mode=self.common_merge, name='fused_em')
else:
fused_emb = ind_emb[0]
return ind_emb + [fused_emb]
def get_input(self, modality):
assert modality in self.input_modalities
for l in self.model.layers:
if l.name == 'enc_' + modality + '_input':
return l.output
return None
def predict_z(self, input_modalities, data, ids):
embeddings = self.get_embeddings(input_modalities)
inputs = [self.get_input(mod) for mod in input_modalities]
partial_model = Model(input=inputs, output=embeddings)
X = [data.select_for_ids(inmod, ids) for inmod in input_modalities]
Z = partial_model.predict(X)
assert len(Z) == len(embeddings)
return Z
def new_decoder_model(self, input_modalities, modality):
if modality in self.output_modalities:
print 'Using trained decoder'
decoder = self.decoders[self.output_modalities.index(modality)]
else:
print 'Creating new decoder'
decoder = self.decoder_maker(modality)
inputs = [Input(shape=(self.latent_dim, None, None)) for i in range(len(input_modalities) + 1)]
outputs = [decoder(inpt) for inpt in inputs]
for outi, out in enumerate(outputs):
out.name = 'em_%d_dec_%s' % (outi, modality)
out_dict = {decoder.name: mae}
loss_weights = {decoder.name: 1.0}
new_model = Model(input=inputs, output=outputs)
new_model.compile(optimizer=Adam(lr=0.0001), loss=out_dict, loss_weights=loss_weights)
return new_model
def get_partial_model(self, input_modalities, output_modality):
assert set(input_modalities).issubset(set(self.input_modalities))
assert output_modality in self.output_modalities
inputs = self.get_inputs(input_modalities)
embeddings = self.get_embeddings(input_modalities)
decoder = self.decoders[self.output_modalities.index(output_modality)]
outputs = get_decoder_outputs([output_modality], [decoder], embeddings)
outputs += self.get_embedding_distance_outputs(embeddings)
model = Model(input=inputs, output=outputs)
return model
def new_encoder_model(self, modality, output_modalities):
if modality in self.input_modalities:
print 'Using trained encoder'
input = self.inputs[self.input_modalities.index(modality)]
lr = self.all_emb[self.input_modalities.index(modality)]
else:
print 'Creating new encoder'
input, lr = self.encoder_maker(modality)
decoders = [self.decoders[self.output_modalities.index(mod)] for mod in output_modalities]
for d in decoders:
d.trainable = False
outputs = get_decoder_outputs(output_modalities, decoders, [lr])
model = Model(input=[input], output=outputs)
model.compile(optimizer=Adam(), loss={d.name: mae for d in decoders},
loss_weights={d.name: 1.0 for d in decoders})
return model
def get_decoder_outputs(output_modalities, decoders, embeddings):
assert len(output_modalities) == len(decoders)
outputs = list()
for di, decode in enumerate(decoders):
for emi, em in enumerate(embeddings):
out_em = decode(em)
name = 'em_' + str(emi) + '_dec_' + output_modalities[di]
l = Lambda(lambda x: x + 0, name=name)(out_em)
l.name = name
outputs.append(l)
print 'making output:', em.type, out_em.type, name
return outputs
def embedding_distance(y_true, y_pred):
return K.var(y_pred, axis=1)
def new_flatten(emb, name=''):
l = Lambda(lambda x: K.batch_flatten(x))(emb)
l = Lambda(lambda x: K.expand_dims(x, axis=1), name=name)(l)
return l
def mae(y_true, y_pred):
return K.mean(K.abs(y_pred - y_true))
def var(embeddings):
emb = embeddings[0]
shape = (emb.shape[1], emb.shape[2], emb.shape[3])
sz = shape[0] * shape[1] * shape[2]
flat_embs = [K.reshape(emb, (emb.shape[0], 1, sz)) for emb in embeddings]
emb_var = K.var(K.concatenate(flat_embs, axis=1), axis=1, keepdims=True)
return K.reshape(emb_var, embeddings[0].shape)
def zeros_for_var(emb):
l = Lambda(lambda x: K.zeros_like(x))(emb)
return l
def STMerge(to_merge):
theta, input = to_merge
theta.reshape((input.shape[0], 2, 3))
return SpatialTransformer._transform(theta, input, 1)
def tpn_maker(input_shape):
# initial weights
b = np.zeros((2, 3), dtype='float32')
b[0, 0] = 1
b[1, 1] = 1
W = np.zeros((50, 6), dtype='float32')
weights = [W, b.flatten()]
# input_shape = (1, 112, 80)
target_input = Input(shape=input_shape)
input = Input(shape=input_shape)
stacked = merge([target_input, input], mode='concat')
mp1 = MaxPooling2D(pool_size=(2, 2))(stacked)
conv1 = Conv2D(8, 5)(mp1)
mp2 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(8, 5)(mp2)
mp3 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D(8, 5)(mp3)
flt = Flatten()(conv3)
d50 = Dense(50)(flt)
act = Activation('relu')(d50)
theta = Dense(6, weights=weights)(act)
model = Model(input=[target_input, input], output=theta)
return model