-
Notifications
You must be signed in to change notification settings - Fork 0
/
models_fusion_cls_imgaug.py
507 lines (395 loc) · 16.4 KB
/
models_fusion_cls_imgaug.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
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
"""
this is for classification metastasis or not
"""
import os
from skimage import io
import tensorflow as tf
from utils import *
import utils
import keras.backend as K
import numpy as np
import math
from keras.preprocessing.image import ImageDataGenerator
from resnet_101 import resnet101_model
from resnet_50 import resnet50_model
from test_net import testnet_model
from resnet18 import ResnetBuilder
from snet import snet
#from snet_true_fusion import snet
#from snet_true import snet
from test_snet import test_snet
from grad_cam import *
from imgaug import augmenters as iaa
datagen = ImageDataGenerator(
#rotation_range=30,
#width_shift_range=0.2,
#height_shift_range=0.2,
#zoom_range = 0.2,
#horizontal_flip=False,
#vertical_flip=False
)
# shear_range=0.5,
seq = iaa.Sequential([
iaa.Crop(percent=(0, 0.1)),
iaa.Fliplr(0.5),
iaa.Flipud(0.5),
iaa.Add((-0.1, 0.1)),
iaa.Multiply((0.9, 1.1)),
iaa.OneOf([
iaa.Affine(scale=(0.9, 1.1)),
iaa.Affine(translate_percent={"x": (-0.1, 0.1), "y": (-0.1, 0.1)}),
iaa.Affine(rotate=(-10, 10)),
#iaa.Affine(shear=(-10, 10)),
]),
])
class SurvivalModel:
def __init__(self):
'''
Args
model_builder: a function which returns user defined Keras model
'''
#self.model_builder = resnet101_model
#self.model_builder = resnet50_model
#self.model_builder = testnet_model
self.model_builder = ResnetBuilder.build_resnet_18
#self.model_builder = snet
#self.model_builder = test_snet
def fit(self, datasets_train, datasets_val, datasets_test, train_name=None, val_name=None, test_name=None, loss_func='cross_entropy', epochs=500, lr=0.001, mode='merge', batch_size = 8):
'''
Train a deep survival model
Args
datasets_train: training datasets, a list of (X, time, event) tuples
datasets_val: validation datasets, a list of (X, time, event) tuples
loss_func: loss function to approximate concordance index, {'hinge', 'log', 'cox'}
epochs: number of epochs to train
lr: learning rate
mode: if mode=='merge', merge datasets before training
if mode='decentralize', treat each dataset as a mini-batch
batch_size: only effective for 'merge' mode
'''
self.datasets_train = datasets_train
self.datasets_val = datasets_val
self.datasets_test = datasets_test
self.loss_func = loss_func
self.epochs = epochs
self.lr = lr
self.batch_size = batch_size
self.train_name = train_name
self.val_name = val_name
self.test_name = test_name
## build a tensorflow graph to define loss function
self.__build_graph()
## train the model
if mode=='train':
self.__train_merge()
elif mode=='infer':
self.__infer()
elif mode=='vis':
self.__vis()
elif mode=='vis_cam':
self.__vis_cam()
def __build_graph(self):
'''
Build a tensorflow graph. Call this within self.fit()
'''
input_shape = self.datasets_train[0][0].shape[1:]
n_classes = 2
#print (input_shape[0])
#raise
with tf.name_scope('input'):
X = tf.placeholder(dtype=tf.float32, shape=(None, )+input_shape, name='X')
Y = tf.placeholder(dtype=tf.float32, shape=[None, n_classes], name='Y')
with tf.name_scope('model'):
#self.model = self.model_builder(input_shape[0], input_shape[0])
self.model = self.model_builder((12, 160, 160), 2)
#self.model = self.model_builder()
with tf.name_scope('output'):
score = tf.identity(self.model(X), name='score')
with tf.name_scope('metric'):
acc, correct_prediction = self.__calc_acc(score, Y)
if self.loss_func=='cross_entropy':
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=score, labels=Y))
with tf.name_scope('train'):
num_epoch = tf.Variable(0, name='global_step', trainable=False)
boundaries = [5, 25, 70]
#learing_rates = [0.0001, 0.0005, 0.00001, 0.00005]
#learing_rates = [0.0001, 0.0001, 0.0001, 0.0001]
learing_rates = [0.01, 0.001, 0.001, 0.001]
learing_rate = tf.train.piecewise_constant(num_epoch, boundaries=boundaries, values=learing_rates)
optimizer = tf.train.AdamOptimizer(learning_rate=learing_rate)
#optimizer = tf.train.GradientDescentOptimizer(learning_rate=learing_rate)
#optimizer = tf.train.MomentumOptimizer(learning_rate=learing_rate, momentum=0.9)
#optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
train_op = optimizer.minimize(loss, name='train_op')
## save the tensors and ops so that we can use them later
self.__X = X
self.__Y = Y
self.__score = score
self.__acc = acc
self.__correct_prediction = correct_prediction
self.__loss = loss
self.__train_op = train_op
def __calc_acc(self, score, Y):
correct_prediction = tf.equal(tf.argmax(score, 1), tf.argmax(Y, 1))
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return acc, correct_prediction
def __infer(self):
weights_path = "checkpoints/my_fusion_model_weights_58.h5"
self.__sess = tf.Session()
K.set_session(self.__sess)
self.__sess.run(tf.global_variables_initializer())
self.model.load_weights(weights_path, by_name=True)
self.__print_loss_ci_yh_infer(self.datasets_train, self.train_name, 'train')
self.__print_loss_ci_yh_infer(self.datasets_val, self.val_name, 'val')
self.__print_loss_ci_yh_infer(self.datasets_test, self.test_name, 'test')
def __vis(self):
weights_path = "checkpoints/my_model_weights_35.h5"
self.__sess = tf.Session()
K.set_session(self.__sess)
self.__sess.run(tf.global_variables_initializer())
self.model.load_weights(weights_path, by_name=True)
self.__vis_layer(self.datasets_train, self.train_name, 'train')
def __vis_cam(self):
weights_path = "checkpoints/my_model_weights_113.h5"
#self.__sess = tf.Session()
#K.set_session(self.__sess)
#self.__sess.run(tf.global_variables_initializer())
self.model.load_weights(weights_path, by_name=True)
self.__vis_layer_cam(self.datasets_train, self.train_name, 'train')
def __train_merge(self):
'''
Merge training datasets into a single dataset. Sample mini-batches from the merged dataset for training
'''
#weights_path = 'checkpoints/resnet101_weights_tf.h5'
#weights_path = 'checkpoints/resnet50_weights_tf_dim_ordering_tf_kernels.h5'
## Merge training datasets
#X_train, Y_train = utils.combine_datasets(self.datasets_train)
X, Y = zip(*self.datasets_train)
#print (len(X))
#print ('*'*20)
#print (len(Y))
X_train = np.concatenate(X, axis=0)
Y_train = np.concatenate(Y, axis=0)
#print (len(X_train))
#print ('*'*20)
#print (len(Y_train))
assert len(X_train) == len(Y_train), 'X_train, Y_train len are not equal!!!'
## get training datasets ___ by heng
#X_train, time_train, event_train = utils.get_datasets(self.datasets_train)
## To fetch mini-batches
#next_batch, num_batches = utils.batch_factory(X_train, time_train, event_train, self.batch_size)
## start training
self.__sess = tf.Session()
K.set_session(self.__sess)
self.__sess.run(tf.global_variables_initializer())
#self.__sess.run(tf.truncated_normal_initializer())
#self.model.load_weights(weights_path, by_name=True)
file_train = open(f"train_log.txt","w")
file_val = open(f"val_log.txt","w")
file_test = open(f"test_log.txt","w")
print (f'pre epoch train log:')
acc_train, loss_train = self.__print_loss_acc_yh(self.datasets_train)
acc_val, loss_val = self.__print_loss_acc_yh(self.datasets_val)
acc_test, loss_test = self.__print_loss_acc_yh(self.datasets_test)
file_train.write(str(acc_train)+" "+str(loss_train)+"\n")
file_val.write(str(acc_val)+" "+str(loss_val)+"\n")
file_test.write(str(acc_test)+" "+str(loss_test)+"\n")
acc = 0
for epoch in range(self.epochs):
#next_batch, num_batches = utils.batch_factory(X_train, time_train, event_train, self.batch_size)
#for _ in range(num_batches):
# X_batch, time_batch, event_batch = next_batch()
#time_event = np.hstack((time_train[...,None], event_train[...,None]))
batches = 0
for X_batch, Y_batch in datagen.flow(X_train, Y_train, batch_size=self.batch_size):
batches += 1
if batches >= X_train.shape[0] // self.batch_size:
break
X_batch = seq.augment_images(X_batch)
#self.__sess.run(self.__train_op, feed_dict={self.__X: X_batch, self.__Y: Y_batch, K.learning_phase(): 1})
print (self.__sess.run([self.__train_op, self.__acc], feed_dict={self.__X: X_batch, self.__Y: Y_batch, K.learning_phase(): 1}))
#print (self.__sess.run([self.__score, self.__loss, self.__ci], feed_dict={self.__X: X_batch, self.__time: time_batch, self.__event: event_batch, K.learning_phase(): 1}))
print('-'*20 + 'Epoch: {0}'.format(epoch) + '-'*20)
print (f'epoch {epoch} train log:')
acc_train, loss_train = self.__print_loss_acc_yh(self.datasets_train)
file_train.write(str(acc_train)+" "+str(loss_train)+"\n")
print (f'epoch {epoch} val log:')
acc_val, loss_val = self.__print_loss_acc_yh(self.datasets_val)
file_val.write(str(acc_val)+" "+str(loss_val)+"\n")
#if val_ci > ci:
# ci = val_ci
# self.model.save_weights(f'my_model_weights_{epoch}.h5')
# print ("model saved!")
print (f'epoch {epoch} test log:')
acc_test, loss_test = self.__print_loss_acc_yh(self.datasets_test)
file_test.write(str(acc_test)+" "+str(loss_test)+"\n")
#if ci_test > ci:
#if acc_test > 0.62:
# acc = acc_test
#self.model.save_weights(f'my_fusion_model_weights_{epoch}.h5')
# print ("model saved!")
file_train.close()
file_val.close()
file_test.close()
def __print_loss_acc_yh(self, datasets):
'''
This compute loss and ci on the whole dataset. It is reasonable.
'''
X, Y = zip(*datasets)
X = np.concatenate(X, axis=0)
Y = np.concatenate(Y, axis=0)
assert len(X) == len(Y), 'X, Y len are not equal!!!'
loss_sum = 0
acc_sum = 0
count = 0
bs = 16
for i in range(len(X)//bs):
X_batch, Y_batch = X[bs*i:bs*(i+1),...], Y[bs*i:bs*(i+1),...]
score_i, Y_i = self.__sess.run([self.__score, self.__Y], feed_dict = {self.__X: X_batch, self.__Y: Y_batch, K.learning_phase(): 1})
#print (score_i)
if i == 0:
socre_all = score_i
Y_all = Y_i
else:
socre_all = np.concatenate((socre_all, score_i), axis=0)
Y_all = np.concatenate((Y_all, Y_i), axis=0)
i += 1
X_batch, Y_batch = X[bs*i:len(X),...], Y[bs*i:len(Y),...]
score_i, Y_i = self.__sess.run([self.__score, self.__Y], feed_dict = {self.__X: X_batch, self.__Y: Y_batch, K.learning_phase(): 1})
socre_all = np.concatenate((socre_all, score_i), axis=0)
Y_all = np.concatenate((Y_all, Y_i), axis=0)
#print (socre_all.shape, Y_all.shape)
#print (socre_all, Y_all)
loss, acc = self.__sess.run([self.__loss, self.__acc], feed_dict={self.__score: socre_all, self.__Y: Y_all, K.learning_phase(): 1})
#print (loss, ci)
#raise
#if not (math.isnan(loss_i) or math.isnan(ci_i) or math.isinf(loss_i) or math.isinf(ci_i)):
# loss_sum += loss_i*(len(X)-bs*i)
# ci_sum += ci_i*(len(X)-bs*i)
# count += (len(X)-bs*i)
## print them
#print('loss={0}'.format(round(loss, 3)))
#print('ci={0}'.format(round(ci, 3)))
print(f'loss={loss}')
print(f'acc={acc}')
return acc, loss
def __print_loss_ci_yh_infer(self, datasets, dataset_name, flag = 'train'):
'''
This compute loss and ci on the whole dataset. It is reasonable.
'''
## losses and c-indices on traning
#loss_train = np.zeros(len(self.datasets_train))
#ci_train = np.zeros(len(self.datasets_train))
X, time, event = zip(*datasets)
X = np.concatenate(X, axis=0)
time = np.concatenate(time, axis=0)
event = np.concatenate(event, axis=0)
assert len(X) == len(time) == len(event), print ('X, time, event len are not equal!!!')
loss_sum = 0
ci_sum = 0
count = 0
bs = 16
for i in range(len(X)//bs):
X_batch, time_batch, event_batch = X[bs*i:bs*(i+1),...], time[bs*i:bs*(i+1),...], event[bs*i:bs*(i+1),...]
score_i, time_i, event_i = self.__sess.run([self.__score, self.__time, self.__event], feed_dict = {self.__X: X_batch, self.__time: time_batch, self.__event: event_batch, K.learning_phase(): 1})
if i == 0:
socre_all = score_i
time_all = time_i
event_all = event_i
else:
socre_all = np.concatenate((socre_all, score_i), axis=0)
time_all = np.concatenate((time_all, time_i), axis=0)
event_all = np.concatenate((event_all, event_i), axis=0)
#loss_i, ci_i = self.evaluate(X_batch, time_batch, event_batch)
#print ('loss_i, ci_i: ',loss_i, ci_i)
#if math.isnan(loss_i) or math.isnan(ci_i) or math.isinf(loss_i) or math.isinf(ci_i):
# print ('here')
# continue
#loss_sum += loss_i*bs
#ci_sum += ci_i*bs
#count += bs
#print (loss_i, ci_i, math.isnan(loss_i), math.isnan(ci_i))
#print (X_batch.shape, time_batch.shape, event_batch.shape)
#print (loss_sum, ci_sum, count)
i += 1
X_batch, time_batch, event_batch = X[bs*i:len(X),...], time[bs*i:len(time),...], event[bs*i:len(event),...]
#loss_i, ci_i = self.evaluate(X_batch, time_batch, event_batch)
score_i, time_i, event_i = self.__sess.run([self.__score, self.__time, self.__event], feed_dict = {self.__X: X_batch, self.__time: time_batch, self.__event: event_batch, K.learning_phase(): 1})
socre_all = np.concatenate((socre_all, score_i), axis=0)
time_all = np.concatenate((time_all, time_i), axis=0)
event_all = np.concatenate((event_all, event_i), axis=0)
#print (socre_all.shape, time_all.shape, event_all.shape)
#print (len(socre_all), len(time_all), len(event_all))
cal_ci(socre_all, time_all, event_all, dataset_name, flag)
loss, ci = self.__sess.run([self.__loss, self.__ci], feed_dict={self.__score: socre_all, self.__time: time_all, self.__event: event_all, K.learning_phase(): 1})
#print (loss, ci)
#raise
#if not (math.isnan(loss_i) or math.isnan(ci_i) or math.isinf(loss_i) or math.isinf(ci_i)):
# loss_sum += loss_i*(len(X)-bs*i)
# ci_sum += ci_i*(len(X)-bs*i)
# count += (len(X)-bs*i)
## print them
#print('loss={0}'.format(round(loss, 3)))
#print('ci={0}'.format(round(ci, 3)))
print(f'loss={loss}')
print(f'ci={ci}')
print()
return ci
def __vis_layer(self, datasets, dataset_name, flag = 'train'):
X, time, event = zip(*datasets)
X = np.concatenate(X, axis=0)
time = np.concatenate(time, axis=0)
event = np.concatenate(event, axis=0)
assert len(X) == len(time) == len(event), print ('X, time, event len are not equal!!!')
index = 12
x_in = X[index]
vis_out_dir = f'vis_{index}_layer'
os.makedirs(vis_out_dir, exist_ok=True)
print (x_in.shape)
#print (x_in.max(), x_in.min())
x_in = (x_in-np.min(x_in))/(np.max(x_in)-np.min(x_in))
io.imsave(os.path.join(vis_out_dir, f'ori.jpg'), np.squeeze(x_in))
image_arr=np.reshape(x_in, (-1,160,160,1))
layer = K.function([self.model.layers[0].input], [self.model.layers[3].output])
f1 = layer([image_arr])[0]
re = np.squeeze(np.transpose(f1, (0,3,1,2)))
print (re.shape)
for i in range(re.shape[0]):
temp_im = re[i]
temp_im = (temp_im-np.min(temp_im))/(np.max(temp_im)-np.min(temp_im))
io.imsave(os.path.join(vis_out_dir, f'layer2_{i}.jpg'), temp_im)
return
def __vis_layer_cam(self, datasets, dataset_name, flag = 'train'):
X, time, event = zip(*datasets)
X = np.concatenate(X, axis=0)
time = np.concatenate(time, axis=0)
event = np.concatenate(event, axis=0)
assert len(X) == len(time) == len(event), print ('X, time, event len are not equal!!!')
#index = 1
for index in range(0, len(X)):
x_in = X[index]
name = dataset_name[index]
print (name, index)
raise
vis_out_dir = f'vis_layer_cam'
os.makedirs(vis_out_dir, exist_ok=True)
print (x_in.shape)
print (x_in.max(), x_in.min())
#x_in = (x_in-np.min(x_in))/(np.max(x_in)-np.min(x_in))
io.imsave(os.path.join(vis_out_dir, f'ori_{index}.jpg'), x_in)
image_arr=np.reshape(x_in, (-1,160,160,3))
predictions = self.model.predict(image_arr)
predicted_class = np.argmax(predictions)
print (predictions, predicted_class)
#for layer in self.model.layers:
#print (layer.name)
cam, heatmap = grad_cam(self.model, image_arr, predicted_class, "conv2d_17")
cv2.imwrite(os.path.join(vis_out_dir, f"gradcam_{index}.jpg"), cam)
register_gradient()
guided_model = modify_backprop(self.model, 'GuidedBackProp')
saliency_fn = compile_saliency_function(guided_model)
saliency = saliency_fn([image_arr, 0])
gradcam = saliency[0] * heatmap[..., np.newaxis]
cv2.imwrite(os.path.join(vis_out_dir, f"guided_gradcam_{index}.jpg"), deprocess_image(gradcam))
raise
return