-
Notifications
You must be signed in to change notification settings - Fork 21
/
Copy pathCASED_train.py
313 lines (252 loc) · 13.8 KB
/
CASED_train.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
import time
from heapq import nlargest
from random import uniform
from ops import *
from utils import *
class CASED(object) :
def __init__(self, sess, epoch, batch_size, test_batch_size, num_gpu, checkpoint_dir, result_dir, log_dir):
self.sess = sess
self.dataset_name = 'LUNA16'
self.checkpoint_dir = checkpoint_dir
self.result_dir = result_dir
self.log_dir = log_dir
self.epoch = epoch
self.batch_size = batch_size
self.test_batch_size = test_batch_size
self.predictor_batch_size = batch_size * 2
self.model_name = "CASED" # name for checkpoint
self.num_gpu = num_gpu
self.total_subset = 10
self.x = 68
self.y = 68
self.z = 68
self.c_dim = 1
self.y_dim = 2 # nodule ? or non_nodule ?
self.out_dim = 8
self.weight_decay = 1e-4
self.lr_decay = 1.0 # not decay
self.learning_rate = 0.01
self.momentum = 0.9
self.M = 10
self.K_fold = True
def cased_network(self, x, reuse=False, scope='CASED_NETWORK'):
with tf.variable_scope(scope, reuse=reuse) :
# print(np.shape(x))
x = conv_layer(x, channels=32, kernel=3, stride=1, layer_name='conv1')
up_conv1 = conv_layer(x, channels=32, kernel=3, stride=1, layer_name='up_conv1')
x = max_pooling(up_conv1)
x = conv_layer(x, channels=64, kernel=3, stride=1, layer_name='conv2')
up_conv2 = conv_layer(x, channels=64, kernel=3, stride=1, layer_name='up_conv2')
x = max_pooling(up_conv2)
x = conv_layer(x, channels=128, kernel=3, stride=1, layer_name='conv3')
up_conv3 = conv_layer(x, channels=128, kernel=3, stride=1, layer_name='up_conv3')
x = max_pooling(up_conv3)
x = conv_layer(x, channels=256, kernel=3, stride=1, layer_name='conv4')
x = conv_layer(x, channels=128, kernel=3, stride=1, layer_name='conv5')
x = deconv_layer(x, channels=128, kernel=4, stride=2, layer_name='deconv1')
x = copy_crop(crop_layer=up_conv3, in_layer=x)
x = conv_layer(x, channels=128, kernel=1, stride=1, layer_name='conv6')
x = conv_layer(x, channels=64, kernel=1, stride=1, layer_name='conv7')
x = deconv_layer(x, channels=64, kernel=4, stride=2, layer_name='deconv2')
x = copy_crop(crop_layer=up_conv2, in_layer=x)
x = conv_layer(x, channels=64, kernel=1, stride=1, layer_name='conv8')
x = conv_layer(x, channels=32, kernel=1, stride=1, layer_name='conv9')
x = deconv_layer(x, channels=32, kernel=4, stride=2, layer_name='deconv3')
x = copy_crop(crop_layer=up_conv1, in_layer=x)
x = conv_layer(x, channels=32, kernel=1, stride=1, layer_name='conv10')
x = conv_layer(x, channels=32, kernel=1, stride=1, layer_name='conv11')
logits = conv_layer(x, channels=2, kernel=1, stride=1, activation=None, layer_name='conv12')
x = softmax(logits)
return logits, x
def build_model(self):
patch_dims = [None, None, None, self.c_dim]
bs = self.batch_size
p_bs = self.predictor_batch_size * self.num_gpu
y_dims = [None, None, None, self.y_dim]
self.decay_lr = tf.placeholder(tf.float32, name='learning_rate')
""" Graph Input """
# images
self.inputs = tf.placeholder(tf.float32, [bs] + patch_dims, name='patch')
self.p_inputs = tf.placeholder(tf.float32, [p_bs] + patch_dims, name='p_patch')
# labels
self.y = tf.placeholder(tf.float32, [bs] + y_dims, name='y') # for loss
self.p_y = tf.placeholder(tf.float32, [p_bs] + y_dims, name='p_y')
self.logits, self.softmax_logits = self.cased_network(self.inputs)
self.P_logits, self.P_softmax_logits = self.cased_network(self.inputs, reuse=True)
""" Predictor Loss """
self.x_dict = dict()
for gpu_i in range(self.num_gpu) :
with tf.device('/gpu:%d' % gpu_i) :
index_start = self.predictor_batch_size * gpu_i
index_end = self.predictor_batch_size * (gpu_i + 1)
self.predictor_logits, _ = self.cased_network(self.p_inputs[index_start : index_end], reuse=True)
self.P_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=self.p_y[index_start : index_end],
logits=self.predictor_logits), axis=[1, 2, 3])
self.x_dict.update(
{index_start + i: self.P_loss[i] for i in range(self.predictor_batch_size)}
)
""" Loss function """
self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=self.y, logits=self.logits))
self.l2_loss = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
self.loss = self.loss + self.l2_loss*self.weight_decay
self.correct_prediction = tf.equal(tf.argmax(self.softmax_logits, -1), tf.argmax(self.y, -1))
self.P_correct_prediction = tf.equal(tf.argmax(self.P_softmax_logits, -1), tf.argmax(self.y, -1))
self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction, tf.float32))
self.P_accuracy = tf.reduce_mean(tf.cast(self.P_correct_prediction, tf.float32))
self.sensitivity, self.fp_rate = sensitivity(labels=self.y, logits=self.softmax_logits)
self.P_sensitivity, self.P_fp_rate = sensitivity(labels=self.y, logits=self.P_softmax_logits)
self.optim = tf.train.MomentumOptimizer(learning_rate=self.decay_lr, momentum=self.momentum, use_nesterov=True).minimize(self.loss)
""" Summary """
c_lr = tf.summary.scalar('cosine_lr', self.decay_lr)
c_loss = tf.summary.scalar('loss', self.loss)
c_acc = tf.summary.scalar('acc', self.accuracy)
c_recall = tf.summary.scalar('sensitivity', self.sensitivity)
c_fp = tf.summary.scalar('false_positive', self.fp_rate)
self.c_sum = tf.summary.merge([c_lr, c_loss, c_acc, c_recall, c_fp])
def train(self):
# initialize all variables
tf.global_variables_initializer().run()
# saver to save model
self.saver = tf.train.Saver()
# summary writer
self.writer = tf.summary.FileWriter(self.log_dir + '/' + self.model_name, self.sess.graph)
# restore check-point if it exits
could_load, checkpoint_counter = self.load(self.checkpoint_dir)
if could_load:
if self.K_fold :
start_epoch = (int)(checkpoint_counter / (self.total_subset - 1))
start_sub_n = checkpoint_counter - start_epoch * (self.total_subset - 1) + 1
else :
start_epoch = (int)(checkpoint_counter / (self.total_subset))
start_sub_n = checkpoint_counter - start_epoch * (self.total_subset)
counter = checkpoint_counter
print(" [*] Load SUCCESS")
else:
start_epoch = 0
start_sub_n = 0
counter = 0
print(" [!] Load failed...")
# loop for epoch
start_time = time.time()
count_temp = 0
# 10 counter = 1 epoch
for epoch in range(start_epoch, self.epoch):
validation_sub_n = (epoch % self.total_subset)
print('validation sub_n : {}'.format(validation_sub_n))
for sub_n in range(start_sub_n, self.total_subset) :
# K fold cross validation ...
if sub_n == validation_sub_n :
continue
train_acc = 0.0
train_recall = 0.0
train_fp = 0.0
nan_num = 0
prob = 1.0
train_lr = self.learning_rate
nodule_patch, all_patch, nodule_y, all_y = prepare_data(sub_n)
M = len(all_patch)
num_batches = M // self.batch_size
print('finish prepare data : ', M)
predict_batch = self.predictor_batch_size * self.num_gpu
total_predict_index = M // predict_batch
for idx in range(num_batches):
if idx == int((num_batches * 0.5)) :
train_lr = train_lr * self.lr_decay
print("*** now learning rate : {} ***\n".format(train_lr))
# train_lr = Snapshot(t=idx, T=num_batches, M=self.M, alpha_zero=self.learning_rate)
each_time = time.time()
p = uniform(0,1)
print('probability M : ', prob)
print('condition P : ', p)
if p <= prob :
random_index = np.random.choice(len(nodule_patch), size=self.batch_size, replace=False)
batch_patch = nodule_patch[random_index]
batch_y = nodule_y[random_index]
else :
predict_dict = dict()
for p_idx in range(total_predict_index) :
result = dict()
batch_patch = all_patch[predict_batch*p_idx : predict_batch*(p_idx+1)]
batch_y = all_y[predict_batch*p_idx : predict_batch*(p_idx+1)]
predictor_feed_dict = {
self.p_inputs: batch_patch, self.p_y : batch_y
}
temp_x = self.sess.run(self.x_dict, feed_dict=predictor_feed_dict)
if p_idx != 0 :
for k,v in temp_x.items() :
new_k = k + predict_batch * p_idx
result[new_k] = v
else :
for k,v in temp_x.items() :
result[k] = v
predict_dict.update(result)
index = nlargest(self.batch_size, predict_dict, key=predict_dict.get)
predict_dict = {s_idx: predict_dict[s_idx] for s_idx in index}
g_r_index = list(predict_dict.keys())
batch_patch = all_patch[g_r_index]
batch_y = all_y[g_r_index]
prob *= pow(1/M, 1/num_batches)
train_feed_dict = {
self.inputs: batch_patch, self.y : batch_y,
self.decay_lr : train_lr
}
_, summary_str_c, c_loss, c_acc, c_recall, c_fp = self.sess.run(
[self.optim, self.c_sum, self.loss, self.accuracy, self.sensitivity, self.fp_rate],
feed_dict=train_feed_dict)
self.writer.add_summary(summary_str_c, count_temp)
count_temp += 1
if np.isnan(c_recall) :
train_acc += c_acc
train_fp += c_fp
# train_recall += 0
nan_num += 1
else :
train_acc += c_acc
train_fp += c_fp
train_recall += c_recall
# display training status
print("Epoch: [%2d], Sub_n: [%2d], [%4d/%4d] time: %4.4f, each_time: %4.4f, c_loss: %.8f, c_acc: %.4f, c_recall: %.4f, c_fp: %.4f" \
% (epoch, sub_n, idx, num_batches, time.time() - start_time, time.time() - each_time, c_loss, c_acc, c_recall, c_fp))
train_acc /= num_batches
train_recall /= (num_batches - nan_num)
train_fp /= num_batches
summary_train = tf.Summary(value=[tf.Summary.Value(tag='train_accuracy', simple_value=train_acc),
tf.Summary.Value(tag='train_recall', simple_value=train_recall),
tf.Summary.Value(tag='train_fp', simple_value=train_fp)])
self.writer.add_summary(summary_train, counter)
line = "Epoch: [%2d], Sub_n: [%2d], train_acc: %.4f, train_recall: %.4f, train_fp: %.4f\n" % (epoch, sub_n, train_acc, train_recall, train_fp)
print(line)
with open(os.path.join(self.result_dir, 'train_logs.txt'), 'a') as f:
f.write(line)
# save model
counter += 1
self.save(self.checkpoint_dir, counter)
del nodule_patch
del nodule_y
del all_patch
del all_y
start_sub_n = 0
@property
def model_dir(self):
return "{}_{}_{}".format(
self.model_name, self.dataset_name,
self.batch_size,)
def save(self, checkpoint_dir, step):
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir, self.model_name)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
self.saver.save(self.sess, os.path.join(checkpoint_dir, self.model_name + '.model'), global_step=step)
def load(self, checkpoint_dir):
import re
print(" [*] Reading checkpoints...")
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir, self.model_name)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
counter = int(next(re.finditer("(\d+)(?!.*\d)", ckpt_name)).group(0))
print(" [*] Success to read {}".format(ckpt_name))
return True, counter
else:
print(" [*] Failed to find a checkpoint")
return False, 0