-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_sgmt.py
287 lines (252 loc) · 12.2 KB
/
train_sgmt.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
import os
import fnmatch
import numpy as np
import matplotlib
matplotlib.use('TkAgg')
import matplotlib.pyplot as plt
from keras.utils import np_utils
from keras import backend as K
from keras.optimizers import Adam, SGD
from keras.callbacks import ModelCheckpoint
from keras.preprocessing.image import ImageDataGenerator
from keras.models import load_model
import ISIC2018_dataset as ISIC
import models
from metrics import dice_loss, jacc_loss, jacc_coef, dice_jacc_mean
from isic_utils import random_crop_image, save_pred_images, save_chw_image
import pdb
# Base Environment
project_path = "/home/zichen/segment/"
K.set_image_dim_ordering('th') # Theano dimension ordering: (channels, width, height)
np.random.seed(4)
seed = 1
# Model Settings
pre_model = None #project_path + "weights/{}.h5".format('test')
model_type = 'vgg'
model_filename = project_path + "weights/temp.h5" # Suggest that everytime a new model trained as temp.h5 and be renamed later.
custom_loss = dice_loss
optimizer = Adam(lr=1e-5)
custom_metric = [jacc_coef]
# Training Control
do_stage = 1 #Stage1: only training set is provided.
#Stage2: trainging/validation set are provided.
#Stage3: training/validation/test set are provided.
batch_size = 32
initial_epoch = 0
n_epoch = 500
fc_size = 4000
do_resize = True
do_train = True
do_evaluate = True
do_test = True
height = 128
width = 128
channels = 3
remove_mean_imagenet = False
remove_mean_dataset = False
remove_mean_samplewise = False
rescale_mask = True # Attention: rescale_mask must be true cause in later evaluation, it will multiple 255
datasets_div = 'manu' # 'isic', 'manu' for division type
year = "2017"
# Creat image lists from training/validation/test
if datasets_div == 'isic':
if year is "2017":
tr_folder = project_path + "datasets/2017/ISIC-2017_Training_Data"
tr_mask_folder = project_path + "datasets/2017/ISIC-2017_Training_Part1_GroundTruth"
tr_csv_file = project_path + "datasets/2017/ISIC-2017_Training_Part3_GroundTruth.csv"
val_folder = project_path + "datasets/2017/ISIC-2017_Validation_Data"
val_mask_folder = project_path + "datasets/2017/ISIC-2017_Validation_Part1_GroundTruth"
val_csv_file = project_path + "datasets/2017/ISIC-2017_Validation_Part3_GroundTruth.csv"
te_folder = project_path + "datasets/2017/ISIC-2017_Test_v2_Data"
te_mask_folder = project_path + "datasets/2017/ISIC-2017_Test_v2_Part1_GroundTruth"
te_csv_file = project_path + "datasets/2017/ISIC-2017_Test_v2_Part3_GroundTruth.csv"
resized_image_folder = project_path + "datasets/2017/resized"
if year is "2018":
tr_folder = project_path + "datasets/2018/ISIC-2018_Training_Data"
tr_mask_folder = project_path + "datasets/2018/ISIC-2018_Training_Part1_GroundTruth"
val_folder = project_path + "datasets/2018/ISIC-2018_Validation_Data"
val_mask_folder = project_path + "datasets/2018/ISIC-2018_Validation_Part1_GroundTruth"
te_folder = project_path + "datasets/2018/ISIC-2018_Test_v2_Data"
te_mask_folder = project_path + "datasets/2018/ISIC-2018_Test_v2_Part1_GroundTruth"
resized_image_folder = project_path + "datasets/2018/resized"
elif datasets_div == 'manu':
tr_folder = project_path + "datasets/manual/Training"
tr_mask_folder = project_path + "datasets/manual/Training_GroundTruth"
val_folder = project_path + "datasets/manual/Validation"
val_mask_folder = project_path + "datasets/manual/Validation_GroundTruth"
te_folder = project_path + "datasets/manual/Test"
te_mask_folder = project_path + "datasets/manual/Test_GroundTruth"
resized_image_folder = project_path + "datasets/manual/resized"
pred_mask_folder = project_path + "pred_picture/"
pred_val_mask_folder = pred_mask_folder + "validation"
pred_te_mask_folder = pred_mask_folder + "test"
def myGenerator(image_generator, mask_generator):
while True:
image_gen = next(image_generator)
mask_gen = next(mask_generator)
image_gen = image_gen/127.5 - 1
yield (image_gen, mask_gen)
if pre_model:
model = load_model(pre_model, custom_objects={'dice_loss': dice_loss, 'jacc_coef': jacc_coef})
print "Loaded previous model.\n"
model.compile(optimizer=optimizer, loss={
'conv7': custom_loss,
'fc2': 'categorical_crossentropy',
},
loss_weights = {
'conv7': 1.,
'fc2': 0.,
},
metrics={
'conv7': custom_metric,
#'fc2': 'categorical_accuracy',
})
monitor_metric = 'val_conv7_jacc_coef' #only when multi-output it will be like ._conv7_.
elif model_type == 'unet':
model = models.Unet(height, width, custom_loss=custom_loss, optimizer=optimizer, custom_metrics=custom_metric, fc_size=fc_size, channels=channels)
monitor_metric = 'val_jacc_coef'
elif model_type == 'unet2':
model = models.Unet2(height, width, custom_loss=custom_loss, optimizer=optimizer, custom_metrics=custom_metric, fc_size=fc_size, channels=channels)
monitor_metric = 'val_jacc_coef'
elif model_type =='vgg':
VGG16_WEIGHTS_NOTOP = project_path + 'pretrained_weights/vgg16_notop.h5'
model = models.VGG16(height, width, pretrained=VGG16_WEIGHTS_NOTOP, freeze_pretrained = False, custom_loss = custom_loss, optimizer = optimizer, custom_metrics = custom_metric)
monitor_metric = 'val_jacc_coef'
if do_stage == 1:
split_ratio = [4, 1, 1]
tr_list, val_list, te_list = ISIC.split_isic_train(tr_folder, split_ratio)
val_folder = tr_folder
val_mask_folder = tr_mask_folder
te_folder = tr_folder
te_mask_folder = tr_mask_folder
base_tr_folder = resized_image_folder + "/train_{}_{}".format(height, width)
base_val_folder = resized_image_folder + "/validation_{}_{}".format(height, width)
base_te_folder = resized_image_folder + "/test_{}_{}".format(height, width)
elif do_stage == 2:
# tr and val are same with do_split_train_data= True, test is change to val. This is for second stage.
validation_folder = val_folder
te_folder = val_folder
te_mask_folder = val_mask_folder
te_list = fnmatch.filter(os.listdir(validation_folder), '*.jpg')
base_te_folder = resized_image_folder + "/test_{}_{}".format(height, width)
split_ratio = [7, 1, 0]
tr_list, val_list, false_telist = ISIC.split_isic_train(tr_folder, split_ratio)
val_folder = tr_folder #here change
val_mask_folder = tr_mask_folder
base_tr_folder = resized_image_folder + "/train_{}_{}".format(height, width)
base_val_folder = resized_image_folder + "/validation_{}_{}".format(height, width)
elif do_stage == 3:
tr_list = fnmatch.filter(os.listdir(tr_folder), '*.jpg')
val_list = fnmatch.filter(os.listdir(val_folder), '*.jpg')
te_list = fnmatch.filter(os.listdir(te_folder), '*.jpg')
base_tr_folder = resized_image_folder + "/train_{}_{}".format(height, width)
base_val_folder = resized_image_folder + "/validation_{}_{}".format(height, width)
base_te_folder = resized_image_folder + "/test_{}_{}".format(height, width)
# Check folder which stored the resized images
base_tr_image_folder = os.path.join(base_tr_folder, "image")
base_tr_mask_folder = os.path.join(base_tr_folder, "mask")
base_val_image_folder = os.path.join(base_val_folder, "image")
base_val_mask_folder = os.path.join(base_val_folder, "mask")
base_te_image_folder = os.path.join(base_te_folder, "image")
base_te_mask_folder = os.path.join(base_te_folder, "mask")
# Resize images and restore resized images to new folder
if do_resize:
print("Begin resizing Images...")
if not os.path.exists(base_tr_folder):
ISIC.resize_image(tr_folder, tr_mask_folder, tr_list, base_tr_image_folder,
base_tr_mask_folder, height, width)
if not os.path.exists(base_val_folder):
ISIC.resize_image(val_folder, val_mask_folder, val_list, base_val_image_folder,
base_val_mask_folder, height, width)
if not os.path.exists(base_te_folder):
ISIC.resize_image(te_folder, te_mask_folder, te_list,
base_te_image_folder, base_te_mask_folder, height, width)
[val_image, val_mask] = ISIC.load_image(
base_val_image_folder, base_val_mask_folder,
val_list, height, width, remove_mean_imagenet,
remove_mean_dataset,rescale_mask)
val_image = val_image/127.5 - 1
if do_train:
# Data augmentation and generation
if not os.path.exists(base_tr_folder):
print "No prepared training data"
[tr_image, tr_mask] = ISIC.load_image(
base_tr_image_folder, base_tr_mask_folder,
tr_list, height, width, remove_mean_imagenet,
remove_mean_dataset,rescale_mask)
data_gen_args = dict(
featurewise_center=False,
samplewise_center=remove_mean_samplewise,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=270,
width_shift_range=0.1,
height_shift_range=0.1,
zoom_range=0.2,
channel_shift_range=0,
shear_range=0.2,
fill_mode='reflect',
horizontal_flip=False,
vertical_flip=False,
# preprocessing_function=random_crop_image,
dim_ordering=K.image_dim_ordering())
data_gen_mask_args = dict(
data_gen_args.items() + {'fill_mode':'nearest',
'samplewise_center':False,
'featurewise_center':False,
'featurewise_std_normalization':False,
'samplewise_std_normalization':False,
}.items(),)
image_augment = ImageDataGenerator(**data_gen_args)
mask_augment = ImageDataGenerator(**data_gen_mask_args)
image_augment.fit(tr_image, augment=True, seed=1)
tr_image_generator = image_augment.flow(tr_image, batch_size=batch_size, seed=seed)
tr_mask_generator = mask_augment.flow(tr_mask, batch_size=batch_size, seed=seed)
tr_data_generator = myGenerator(tr_image_generator, tr_mask_generator)
# Check model
model_checkpoint = ModelCheckpoint(model_filename, monitor=monitor_metric,
save_best_only=True, verbose=1)
# Training
n_samples = len(tr_list)
history = model.fit_generator(tr_data_generator,
samples_per_epoch=n_samples,
nb_epoch=n_epoch,
validation_data=(val_image, val_mask),
callbacks=[model_checkpoint],
initial_epoch=initial_epoch)
train = None; train_mask = None
# Evaluate model: evaluate val
if do_evaluate:
model.load_weights(model_filename)
val_pred_mask = model.predict(val_image)
save_pred_images(val_pred_mask, val_list, pred_val_mask_folder)
for pixel_threshold in [0.5]: # np.arange(0.3, 1, 0.05):
val_pred_mask = np.where(val_pred_mask>=pixel_threshold, 1, 0)
val_pred_mask = val_pred_mask * 255
val_pred_mask = val_pred_mask.astype(np.uint8)
print "Validation Prediction Max:{}, Min:{}".format(np.max(val_pred_mask),
np.min(val_pred_mask))
print model.evaluate(val_image, val_mask, batch_size = batch_size, verbose=1)
dice, jacc = dice_jacc_mean(val_mask, val_pred_mask, smooth = 0)
print "Resized val dice coef: {:.4f}".format(dice)
print "Resized val jacc coef: {:.4f}".format(jacc)
if do_test:
[te_image, te_mask] = ISIC.load_image(
base_te_image_folder, base_te_mask_folder,
te_list, height, width, remove_mean_imagenet,
remove_mean_dataset, rescale_mask)
te_image = te_image/127.5 - 1
model.load_weights(model_filename)
te_pred_mask = model.predict(te_image)
save_pred_images(te_pred_mask, te_list, pred_te_mask_folder)
for pixel_threshold in [0.5]: # np.arange(0.3, 1, 0.05):
te_pred_mask = np.where(te_pred_mask>=pixel_threshold, 1, 0)
te_pred_mask = te_pred_mask * 255
te_pred_mask = te_pred_mask.astype(np.uint8)
print "Test Prediction Max:{}, Min:{}".format(np.max(te_pred_mask),
np.min(te_pred_mask))
print model.evaluate(te_image, te_mask, batch_size = batch_size, verbose=1)
dice, jacc = dice_jacc_mean(te_mask, te_pred_mask, smooth = 0)
print "Resized te dice coef: {:.4f}".format(dice)
print "Resized te jacc coef: {:.4f}".format(jacc)