-
Notifications
You must be signed in to change notification settings - Fork 3
/
retinaNN_training.py
279 lines (218 loc) · 11.2 KB
/
retinaNN_training.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
###################################################
#
# Script to:
# - Load the images and extract the patches
# - Define the neural network
# - define the training
#
##################################################
# -------------------------- set gpu using tf ---------------------------
from __future__ import absolute_import, division, print_function, unicode_literals
import tensorflow as tf
gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(gpus[0], True)
# -----------------------------------------------------------------------
import numpy as np
import configparser
from tensorflow.keras import Model
from tensorflow.keras.layers import Input, concatenate, Conv2D, MaxPooling2D, UpSampling2D, Reshape, Dropout, Conv2DTranspose, Multiply
from tensorflow.keras.optimizers import Adam, Nadam
from tensorflow.keras.callbacks import ModelCheckpoint, LearningRateScheduler, CSVLogger, TensorBoard, EarlyStopping
from tensorflow.keras import backend as K
from tensorflow.keras.utils import plot_model
from tensorflow.keras.optimizers import SGD, schedules
from tensorflow.keras.layers import Dense, Lambda, Activation, Permute, Flatten
from tensorflow.keras import layers
from tensorflow.keras.regularizers import l2
from tensorflow.keras.layers import BatchNormalization, LayerNormalization
from tensorflow.keras.losses import categorical_crossentropy
from DropBlock import DropBlock2D, Mish
import sys
sys.path.insert(0, './lib/')
from help_functions import *
#function to obtain data for training/testing (validation)
from extract_patches import get_data_training
def dice_coef(y_true, y_pred, smooth=1):
"""
Dice = (2*|X & Y|)/ (|X|+ |Y|)
= 2*sum(|A*B|)/(sum(A^2)+sum(B^2))
ref: https://arxiv.org/pdf/1606.04797v1.pdf
"""
y_true_f = K.flatten(y_true[:,:,1])
y_pred_f = K.flatten(y_pred[:,:,1])
intersection = K.sum(y_true_f * y_pred_f)
return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
# return (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
def combo_loss(y_true, y_pred):
dice_loss = K.mean(1-dice_coef(y_true, y_pred))
loss = 0.25*dice_loss + 0.75*categorical_crossentropy(y_true, y_pred)
return loss
def hw_flatten(x):
return tf.reshape(x, [K.shape(x)[0],-1,x.shape[-1]])
def ChannelSE(reduction=4):
"""
Squeeze and Excitation block, reimplementation inspired by
https://github.com/Cadene/pretrained-models.pytorch/blob/master/pretrainedmodels/models/senet.py
Args:
reduction: channels squeeze factor
"""
def layer(input_tensor):
# get number of channels/filters
channels = K.int_shape(input_tensor)[3]
x = input_tensor
# squeeze and excitation block in PyTorch style with
# custom global average pooling where keepdims=True
x = Lambda(lambda a: K.mean(a, axis=[1, 2], keepdims=True))(x)
x = Conv2D(channels // reduction, (1, 1), kernel_initializer='he_uniform', data_format='channels_last')(x)
x = Activation('relu')(x)
x = Conv2D(channels, (1, 1), kernel_initializer='he_uniform', data_format='channels_last')(x)
x = Activation('sigmoid')(x)
# apply attention
x = Multiply()([input_tensor, x])
return x
return layer
# Define buildingbolck unit like ResNet
def residual_block(input_tensor, kernel_size, filters, stage, block, add_bn=True,
dropout=False, pool=True, SE=True, reduction=4, GC=False, GP=False):
"""The identity block is the block that has no conv layer at shortcut.
# Arguments
input_tensor: input tensor
kernel_size: default 3, the kernel size of middle conv layer at main path
filters: list of integers, the filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
add_bn: bool. True if you want to add BatchNormalization layer.
# Returns
Output tensor for the block.
"""
filters1, filters2 = filters
groups = 2
radix = 2
bn_axis = 3
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
if pool:
stride = (1, 1)
filtersin = filters1
else:
stride = (2, 2)
filtersin = filters1 // 2
x = input_tensor
if add_bn:
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Conv2D(filters1, (kernel_size, kernel_size), strides=stride,
padding='same', name=conv_name_base + '2a', data_format='channels_last')(x)
x = Activation('relu')(x)
if add_bn:
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
if dropout:
x = DropBlock2D(7, 0.8)(x)
x = Conv2D(filters2, (kernel_size, kernel_size), padding='same',
name=conv_name_base + '2c', data_format='channels_last')(x)
x = Activation('relu')(x)
if SE:
x = ChannelSE(reduction=reduction)(x)
if x.shape != input_tensor.shape:
shortcut = Conv2D(filters2, (1, 1), strides=stride,
padding='same', name=conv_name_base + '1', data_format='channels_last')(input_tensor)
else:
shortcut = input_tensor
x = layers.add([x, shortcut])
return x
# Define the resU-net which has 3 depth
def bseresunet(patch_height,patch_width,n_ch):
width = 1
original_inputs = Input(shape=(patch_height,patch_width,n_ch))
# K.image_data_format() == 'channels_last'
conv1 = residual_block(original_inputs, 3, [32*width, 32*width], stage=1, block='a', dropout=True, GP=False)
conv1 = residual_block(conv1, 3, [32*width, 32*width], stage=1, block='b', dropout=True)
#conv1 = residual_block(conv1, 3, [32*width, 32*width], stage=1, block='c', dropout=True)
#
conv2 = residual_block(conv1, 3, [64*width, 64*width], stage=2, block='a', dropout=True, pool=False)
conv2 = residual_block(conv2, 3, [64*width, 64*width], stage=2, block='b', dropout=True)
#conv2 = residual_block(conv2, 3, [64*width, 64*width], stage=2, block='c', dropout=True)
#
conv3 = residual_block(conv2, 3, [128*width, 128*width], stage=3, block='a', dropout=True, pool=False)
conv3 = residual_block(conv3, 3, [128*width, 128*width], stage=3, block='b', dropout=True)
#conv3 = residual_block(conv3, 3, [128*width, 128*width], stage=3, block='c', dropout=True)
#conv3 = residual_block(conv3, 3, [128*width, 128*width], stage=3, block='d', dropout=True)
up1 = Conv2DTranspose(128*width, (3, 3), activation='relu', padding='same', strides=(2, 2), data_format='channels_last')(conv3)
up1 = concatenate([conv2,up1],axis=3)
conv4 = residual_block(up1, 3, [64*width, 64*width], stage=4, block='a', dropout=True)
conv4 = residual_block(conv4, 3, [64*width, 64*width], stage=4, block='b', dropout=True)
#conv4 = residual_block(conv4, 3, [64*width, 64*width], stage=4, block='c', dropout=True)
#
up2 = Conv2DTranspose(64*width, (3, 3), activation='relu', padding='same', strides=(2, 2), data_format='channels_last')(conv4)
up2 = concatenate([conv1,up2], axis=3)
conv5 = residual_block(up2, 3, [32*width, 32*width], stage=5, block='a', dropout=True)
conv5 = residual_block(conv5, 3, [32*width, 32*width], stage=5, block='b', dropout=True)
#conv5 = residual_block(conv5, 3, [32*width, 32*width], stage=5, block='c', dropout=True)
#
conv6 = Conv2D(2, (1, 1), activation='relu',padding='same', data_format='channels_last')(conv5)
conv6 = Reshape((patch_height*patch_width,2))(conv6)
############
conv7 = Activation('softmax')(conv6)
model = Model(inputs=original_inputs, outputs=conv7)
sgd = SGD(lr=0.02, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss=combo_loss, metrics=['accuracy'])
return model
#========= Load settings from Config file
config = configparser.RawConfigParser()
config.read('configuration.txt')
#patch to the datasets
path_data = config.get('data paths', 'path_local')
#Experiment name
name_experiment = config.get('experiment name', 'name')
#training settings
N_epochs = int(config.get('training settings', 'N_epochs'))
batch_size = int(config.get('training settings', 'batch_size'))
#============ Load the data and divided in patches
patches_imgs_train, patches_masks_train = get_data_training(
DRIVE_train_imgs_original = path_data + config.get('data paths', 'train_imgs_original'),
DRIVE_train_groudTruth = path_data + config.get('data paths', 'train_groundTruth'), #masks
patch_height = int(config.get('data attributes', 'patch_height')),
patch_width = int(config.get('data attributes', 'patch_width')),
N_subimgs = int(config.get('training settings', 'N_subimgs')),
inside_FOV = config.getboolean('training settings', 'inside_FOV') #select the patches only inside the FOV (default == True)
)
#========= Save a sample of what you're feeding to the neural network ==========
N_sample = min(patches_imgs_train.shape[0],40)
visualize(group_images(patches_imgs_train[0:N_sample,:,:,:],5),'./'+name_experiment+'/'+"sample_input_imgs")#.show()
visualize(group_images(patches_masks_train[0:N_sample,:,:,:],5),'./'+name_experiment+'/'+"sample_input_masks")#.show()
#=========== Construct and save the model arcitecture =====
n_ch = patches_imgs_train.shape[3]
patch_height = patches_imgs_train.shape[1]
patch_width = patches_imgs_train.shape[2]
# print()
model = bseresunet(patch_height, patch_width,n_ch) #the U-net model
print ("Check: final output of the network:")
print (model.output_shape)
# plot_model(model, to_file='./'+name_experiment+'/'+name_experiment + '_model.png') #check how the model looks like
json_string = model.to_json()
open('./'+name_experiment+'/'+name_experiment +'_architecture.json', 'w').write(json_string)
#============ Training ==================================
#save at each epoch if the validation decreased
checkpointer = ModelCheckpoint(filepath='./'+name_experiment+'/'+name_experiment +'_best_weights.h5',
verbose=1, monitor='val_loss', mode='auto', save_best_only=True)
def step_decay(epoch):
lrate = 0.02 #the initial learning rate (by default in keras)
if epoch <= 30:
return lrate
elif epoch <= 50:
return lrate*0.5
else:
return lrate*0.25
lrate_drop = LearningRateScheduler(step_decay, verbose=1)
early_stop = EarlyStopping(monitor='val_loss', min_delta=0, patience=20, verbose=1, mode='auto', baseline=None, restore_best_weights=False)
csv_logger = CSVLogger('./test/training.csv')
visualize = TensorBoard(log_dir='./test/logs')
patches_masks_train = masks_Unet(patches_masks_train) #reduce memory consumption
model.fit(patches_imgs_train, patches_masks_train, epochs=N_epochs,
batch_size=batch_size, verbose=2, shuffle=True, validation_split=0.1,
callbacks=[checkpointer, csv_logger, visualize, lrate_drop, early_stop])
#========== Save and test the last model ===================
model.save_weights('./'+name_experiment+'/'+name_experiment +'_last_weights.h5', overwrite=True)
#test the model
# score = model.evaluate(patches_imgs_test, masks_Unet(patches_masks_test), verbose=0)
# print('Test score:', score[0])
# print('Test accuracy:', score[1])