-
Notifications
You must be signed in to change notification settings - Fork 161
/
train_hybrid.py
218 lines (201 loc) · 8.81 KB
/
train_hybrid.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
"""Test ImageNet pretrained DenseNet"""
from __future__ import print_function
import sys
sys.path.insert(0,'Keras-2.0.8')
from multiprocessing.dummy import Pool as ThreadPool
from medpy.io import load
import numpy as np
from keras.optimizers import SGD
from keras.callbacks import ModelCheckpoint
from hybridnet import dense_rnn_net
from denseunet3d import denseunet_3d
import keras.backend as K
import os
import time
from loss import weighted_crossentropy
from skimage.transform import resize
import argparse
import os
K.set_image_dim_ordering('tf')
# global parameters
parser = argparse.ArgumentParser(description='Keras DenseUnet Training')
# data folder
parser.add_argument('-data', type=str, default='data/', help='test images')
parser.add_argument('-save_path', type=str, default='Experiments/')
# other paras
parser.add_argument('-b', type=int, default=1)
parser.add_argument('-input_size', type=int, default=224)
parser.add_argument('-model_weight', type=str, default='./model/model_best.hdf5')
parser.add_argument('-input_cols', type=int, default=8)
parser.add_argument('-arch', type=str, default='')
# data augment
parser.add_argument('-mean', type=int, default=48)
args = parser.parse_args()
thread_num = 14
liverlist = [32,34,38,41,47,87,89,91,105,106,114,115,119]
def load_seq_crop_data_masktumor_try(Parameter_List):
img = Parameter_List[0]
tumor = Parameter_List[1]
lines = Parameter_List[2]
numid = Parameter_List[3]
minindex = Parameter_List[4]
maxindex = Parameter_List[5]
# randomly scale
scale = np.random.uniform(0.8,1.2)
deps = int(args.input_size * scale)
rows = int(args.input_size * scale)
cols = args.input_cols
sed = np.random.randint(1,numid)
cen = lines[sed-1]
cen = np.fromstring(cen, dtype=int, sep=' ')
# print (cen)
a = min(max(minindex[0] + deps/2, cen[0]), maxindex[0]- deps/2-1)
b = min(max(minindex[1] + rows/2, cen[1]), maxindex[1]- rows/2-1)
c = min(max(minindex[2] + cols/2, cen[2]), maxindex[2]- cols/2-1)
cropp_img = img[a - deps / 2:a + deps / 2, b - rows / 2:b + rows / 2,
c - args.input_cols / 2: c + args.input_cols / 2].copy()
cropp_tumor = tumor[a - deps / 2:a + deps / 2, b - rows / 2:b + rows / 2,
c - args.input_cols / 2:c + args.input_cols / 2].copy()
cropp_img -= args.mean
# randomly flipping
flip_num = np.random.randint(0,8)
if flip_num == 1:
cropp_img = np.flipud(cropp_img)
cropp_tumor = np.flipud(cropp_tumor)
elif flip_num == 2:
cropp_img = np.fliplr(cropp_img)
cropp_tumor = np.fliplr(cropp_tumor)
elif flip_num == 3:
cropp_img = np.rot90(cropp_img, k=1, axes=(1, 0))
cropp_tumor = np.rot90(cropp_tumor, k=1, axes=(1, 0))
elif flip_num == 4:
cropp_img = np.rot90(cropp_img, k=3, axes=(1, 0))
cropp_tumor = np.rot90(cropp_tumor, k=3, axes=(1, 0))
elif flip_num == 5:
cropp_img = np.fliplr(cropp_img)
cropp_tumor = np.fliplr(cropp_tumor)
cropp_img = np.rot90(cropp_img, k=1, axes=(1, 0))
cropp_tumor = np.rot90(cropp_tumor, k=1, axes=(1, 0))
elif flip_num == 6:
cropp_img = np.fliplr(cropp_img)
cropp_tumor = np.fliplr(cropp_tumor)
cropp_img = np.rot90(cropp_img, k=3, axes=(1, 0))
cropp_tumor = np.rot90(cropp_tumor, k=3, axes=(1, 0))
elif flip_num == 7:
cropp_img = np.flipud(cropp_img)
cropp_tumor = np.flipud(cropp_tumor)
cropp_img = np.fliplr(cropp_img)
cropp_tumor = np.fliplr(cropp_tumor)
#
cropp_tumor = resize(cropp_tumor, (args.input_size,args.input_size, args.input_cols), order=0, mode='edge', cval=0, clip=True, preserve_range=True)
cropp_img = resize(cropp_img, (args.input_size,args.input_size, args.input_cols), order=3, mode='constant', cval=0, clip=True, preserve_range=True)
return cropp_img, cropp_tumor
def generate_arrays_from_file(batch_size, trainidx, img_list, tumor_list, tumorlines, liverlines, tumoridx, liveridx, minindex_list, maxindex_list):
while 1:
X = np.zeros((batch_size, args.input_size, args.input_size, args.input_cols,1), dtype='float32')
Y = np.zeros((batch_size, args.input_size, args.input_size, args.input_cols,1), dtype='int16')
Parameter_List = []
for idx in xrange(batch_size):
count = np.random.choice(trainidx)
img = img_list[count]
tumor = tumor_list[count]
minindex = minindex_list[count]
maxindex = maxindex_list[count]
num = np.random.randint(0,6)
if num < 3 or (count in liverlist):
lines = liverlines[count]
numid = liveridx[count]
else:
lines = tumorlines[count]
numid = tumoridx[count]
Parameter_List.append([img, tumor, lines, numid, minindex, maxindex])
pool = ThreadPool(thread_num)
result_list = pool.map(load_seq_crop_data_masktumor_try, Parameter_List)
pool.close()
pool.join()
for idx in xrange(len(result_list)):
X[idx, :, :, :, 0] = result_list[idx][0]
Y[idx, :, :, :, 0] = result_list[idx][1]
if np.sum(Y==0)==0:
continue
if np.sum(Y==1)==0:
continue
if np.sum(Y==2)==0:
continue
yield (X,Y)
def train_and_predict(args):
print('-'*30)
print('Creating and compiling model...')
print('-'*30)
if args.arch == "3dpart":
model = denseunet_3d(args)
model_path = "/3dpart_model"
sgd = SGD(lr=1e-3, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss=[weighted_crossentropy])
model.load_weights(args.model_weight, by_name=True, by_gpu=True, two_model=True, by_flag=True)
else:
model = dense_rnn_net(args)
model_path = "/hybrid_model"
sgd = SGD(lr=1e-3, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss=[weighted_crossentropy])
model.load_weights(args.model_weight)
# liver tumor LITS
trainidx = list(range(131))
img_list = []
tumor_list = []
minindex_list = []
maxindex_list = []
tumorlines = []
tumoridx = []
liveridx = []
liverlines = []
for idx in xrange(131):
img, img_header = load(args.data + '/myTrainingData/volume-' + str(idx) + '.nii' )
tumor, tumor_header = load(args.data + '/myTrainingData/segmentation-' + str(idx) + '.nii')
img_list.append(img)
tumor_list.append(tumor)
maxmin = np.loadtxt(args.data+'/myTrainingDataTxt/LiverBox/box_' + str(idx) + '.txt', delimiter=' ')
minindex = maxmin[0:3]
maxindex = maxmin[3:6]
minindex = np.array(minindex, dtype='int')
maxindex = np.array(maxindex, dtype='int')
minindex[0] = max(minindex[0]-3, 0)
minindex[1] = max(minindex[1]-3, 0)
minindex[2] = max(minindex[2]-3, 0)
maxindex[0] = min(img.shape[0], maxindex[0]+3)
maxindex[1] = min(img.shape[1], maxindex[1]+3)
maxindex[2] = min(img.shape[2], maxindex[2]+3)
minindex_list.append(minindex)
maxindex_list.append(maxindex)
f1 = open(args.data+ '/myTrainingDataTxt/TumorPixels/tumor_' + str(idx) + '.txt','r')
tumorline = f1.readlines()
tumorlines.append(tumorline)
tumoridx.append(len(tumorline))
f1.close()
f2 = open(args.data+ '/myTrainingDataTxt/LiverPixels/liver_' + str(idx) + '.txt','r')
liverline = f2.readlines()
liverlines.append(liverline)
liveridx.append(len(liverline))
f2.close()
if not os.path.exists(args.save_path +model_path):
os.mkdir(args.save_path + model_path)
if not os.path.exists(args.save_path + "/history"):
os.mkdir(args.save_path + '/history')
else:
if os.path.exists(args.save_path + "/history/lossbatch.txt"):
os.remove(args.save_path + '/history/lossbatch.txt')
if os.path.exists(args.save_path + "/history/lossepoch.txt"):
os.remove(args.save_path + '/history/lossepoch.txt')
model_checkpoint = ModelCheckpoint(args.save_path + model_path+'/weights.{epoch:02d}-{loss:.2f}.hdf5', monitor='loss', verbose = 1,
save_best_only=False,save_weights_only=False,mode = 'min', period = 1)
print('-'*30)
print('Fitting model......')
print('-'*30)
steps = 27386 / (args.b * 6)
model.fit_generator(generate_arrays_from_file(args.b, trainidx, img_list, tumor_list, tumorlines, liverlines,
tumoridx, liveridx, minindex_list, maxindex_list),
steps_per_epoch=steps,
epochs= 6000, verbose = 1, callbacks = [model_checkpoint], max_queue_size=10, workers=3, use_multiprocessing=True)
print ('Finised Training .......')
if __name__ == '__main__':
train_and_predict(args)