/
liver_project.py
526 lines (424 loc) · 23.4 KB
/
liver_project.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
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
import math
import sys
import os
import re
import shutil
import tensorflow as tf
import numpy as np
import random
import cv2
from dataPrep import DataPrep
from utils import *
from tqdm import tqdm
from timer import timer_dict, timerStats
from default import config, printCfg
from metrics import lp_metrics
import net
#import net_onlyUnet as net
#import ipdb
config.__dict__.update()
config.name = sys.argv[1]
availableGPU = config.gpu
if availableGPU != 'CPU':
if availableGPU == None:
for gpuId in range(4):
if int(os.popen("nvidia-smi -i {} -q --display=MEMORY | grep -m 1 Free | grep -o '[0-9]*'".format(gpuId)).readlines()[0]) < 1000:
continue
availableGPU = gpuId
break
if availableGPU == None:
print "No available GPU device!"
sys.exit(1)
os.environ['CUDA_VISIBLE_DEVICES'] = str(availableGPU)
print "{} \n {}".format(sys.argv[1],printCfg())
print('[i] Runing on GPU: {}'.format(availableGPU))
#-------------------------------------------------------------------------------
def compute_lr(lr_values, lr_boundaries):
with tf.variable_scope('learning_rate'):
global_step = tf.Variable(0, trainable=False, name='global_step')
lr = tf.train.piecewise_constant(global_step, lr_boundaries, lr_values)
return lr, global_step
#-------------------------------------------------------------------------------
def train(train_data_dir, val_data_dir, out_weight_dir):
#---------------------------------------------------------------------------
# Parse the commandline
#---------------------------------------------------------------------------
args = config
snaps_path = os.path.join(out_weight_dir , config.name)
#---------------------------------------------------------------------------
# Find an existing checkpoint
#---------------------------------------------------------------------------
start_epoch = 0
start_idx = 0
checkpoint_file = args.checkpoint_file
if args.continue_training:
if checkpoint_file is None:
print('[!] No checkpoints found, cannot continue!')
return 1
metagraph_file = checkpoint_file + '.meta'
if not os.path.exists(metagraph_file):
print('[!] Cannot find metagraph'.format(metagraph_file))
return 1
step = os.path.basename(checkpoint_file).split('.')[0]
start_epoch = int(step.split('_')[0]) - 1
start_idx = int(step.split('_')[1]) + 1
#---------------------------------------------------------------------------
# Create a project directory
#---------------------------------------------------------------------------
else:
try:
print('[i] Creating directory {}...'.format(snaps_path))
os.makedirs(snaps_path)
except Exception as e:
print('[!] {}'.format( str(e)))
#return 1
print('[i] Starting at epoch: {}'.format( start_epoch+1))
#---------------------------------------------------------------------------
# Configure the training data
#---------------------------------------------------------------------------
print('[i] Configuring the training data...')
try:
dp = DataPrep(train_data_dir, val_data_dir)
print('[i] # training samples: {}'.format(dp.num_train))
print('[i] # validation samples: {}'.format(dp.num_valid))
print('[i] # batch size train: {}'.format(args.batch_size))
except (AttributeError, RuntimeError) as e:
print('[!] Unable to load training data: ' + str(e))
return 1
#---------------------------------------------------------------------------
# Create the network
#---------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
n_train_batches = int(math.ceil(dp.num_train/args.batch_size))
n_valid_batches = int(math.ceil(dp.num_valid/args.batch_size))
lr_values = args.lr_values.split(';')
try:
lr_values = [float(x) for x in lr_values]
except ValueError:
print('[!] Learning rate values must be floats')
sys.exit(1)
lr_boundaries = args.lr_boundaries.split(';')
try:
lr_boundaries = [int(x)*n_train_batches for x in lr_boundaries]
except ValueError:
print('[!] Learning rate boundaries must be ints')
sys.exit(1)
ret = compute_lr(lr_values, lr_boundaries)
learning_rate, global_step = ret
ucrNet = net.UnetCrfRnn(args.batch_size, args.weight_decay, args.bias_decay)
with tf.variable_scope('train_step'):
train_step = tf.train.AdamOptimizer(learning_rate, args.momentum).minimize( \
ucrNet.loss, global_step=global_step, name='train_step')
init = tf.global_variables_initializer()
init_local = tf.local_variables_initializer()
sess.run(init)
sess.run(init_local)
if (start_epoch != 0) or not checkpoint_file is None:
try:
initialize_variables_from_ckpt(sess, checkpoint_file)
#saver = tf.train.Saver()
#saver.restore(sess, checkpoint_file)
except Exception as E:
print E
model_saver = tf.train.Saver(max_to_keep=args.max_snapshots_keep)
if (not checkpoint_file is None) and not args.continue_training:
sess.run([tf.assign(global_step,0)])
initialize_uninitialized_variables(sess)
#-----------------------------------------------------------------------
# Create various helpers
#-----------------------------------------------------------------------
if not os.path.exists(args.logdir):
os.mkdir(args.logdir)
if os.path.exists(os.path.join(args.logdir , config.name)):
shutil.rmtree(os.path.join(args.logdir , config.name))
os.mkdir(os.path.join(args.logdir , config.name))
else:
os.mkdir(os.path.join(args.logdir , config.name))
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(os.path.join(args.logdir , config.name), sess.graph)
#-----------------------------------------------------------------------
# Summaries
#-----------------------------------------------------------------------
restore = start_epoch != 0
#training_metric = MetricsSummary(sess, summary_writer, ['dice', 'precision', 'recall', 'accuracy'], 1)
validation_metric_liv = MetricsSummary(sess, summary_writer, ['diceLiv', 'precisionLiv', 'recallLiv', 'accuracyLiv'], 1)
validation_metric_les = MetricsSummary(sess, summary_writer, ['diceLes', 'precisionLes', 'recallLes', 'accuracyLes'], 1)
training_imgs = ImageSummary(sess, summary_writer, 'training', restore)
validation_imgs = ImageSummary(sess, summary_writer, 'validation', restore)
training_loss = LossSummary(sess, summary_writer, 'training', args.summary_interval)
validation_loss = LossSummary(sess, summary_writer, 'validation', n_valid_batches)
#---------------------------------------------------------------------------
# metrics initilazing
#---------------------------------------------------------------------------
#metricsTrainLiver = lp_metrics()
#metricsTrainLesion = lp_metrics()
metricsTestLiver = lp_metrics()
metricsTestLesion = lp_metrics()
#-----------------------------------------------------------------------
# Get the initial snapshot of the network
#-----------------------------------------------------------------------
#net_summary_ops = net.build_summaries(restore)
#if start_epoch == 0:
# net_summary = sess.run(net_summary_ops)
# summary_writer.add_summary(net_summary, 0)
#summary_writer.flush()
#-----------------------------------------------------------------------
# Cycle through the epoch
#-----------------------------------------------------------------------
print('[i] Training...')
if start_idx == n_train_batches:
start_idx = 0
start_epoch += 1
for e in range(start_epoch, args.epochs):
training_imgs_samples = []
validation_imgs_samples = []
#-------------------------------------------------------------------
# Train
#-------------------------------------------------------------------
randTrainBatchIdx = random.sample(range(args.summary_interval-1), 10)
generator = dp.train_generator(args.batch_size, args.num_workers)
description = '[i] Train {:>2}/{}'.format(e+1, args.epochs)
for idx, (images, gtSegs) in enumerate(tqdm(generator, total=n_train_batches, initial=start_idx, desc=description, unit='batches', leave=False), start=start_idx):
with timer_dict['train']:
[_, loss, segPrediction] = sess.run([train_step, ucrNet.loss, ucrNet.segPrediction], feed_dict={ucrNet.image: images, ucrNet.gtSeg: gtSegs})
training_loss.add(loss)
if idx in randTrainBatchIdx:
if args.batch_size > 1:
randTrainFrameIdx = np.random.randint(0,args.batch_size-1,1)[0]
else:
randTrainFrameIdx = 0
with timer_dict['summary']:
training_imgs_samples.append((np.copy(images[randTrainFrameIdx,:,:,0]), np.copy(segPrediction[randTrainFrameIdx,:,:]), np.copy(gtSegs[randTrainFrameIdx,:,:,:])))
#timerStats()
iteration = int(tf.train.global_step(sess, global_step))
if iteration == 0 or ((iteration % args.summary_interval) != 0 and (iteration % args.val_interval) != 0):
continue
#-------------------------------------------------------------------
# Write summaries
#-------------------------------------------------------------------
training_loss.push(iteration)
summary = sess.run(merged_summary, feed_dict={ucrNet.image: images,
ucrNet.gtSeg: gtSegs})
summary_writer.add_summary(summary, iteration)
training_imgs.push(iteration, training_imgs_samples)
training_imgs_samples = []
summary_writer.flush()
randTrainBatchIdx = random.sample(range(idx+1, idx+args.summary_interval-1), 10)
#-------------------------------------------------------------------
# Validate
#-------------------------------------------------------------------
if (iteration % args.val_interval) != 0:
continue
randValidBatchIdx = random.sample(range(n_valid_batches-1), 10)
generator = dp.valid_generator(args.batch_size, args.num_workers)
description = '[i] Valid {:>2}/{}'.format(e+1, args.epochs)
for idxTest, (images, gtSegs) in enumerate(tqdm(generator, total=n_valid_batches, desc=description, unit='batches', leave=False)):
[loss, segPrediction] = sess.run([ucrNet.loss, ucrNet.segPrediction], feed_dict={ucrNet.image: images, ucrNet.gtSeg: gtSegs})
validation_loss.add(loss)
liverPrediction = np.zeros((segPrediction.shape), dtype=np.float32)
liverPrediction[segPrediction == 1] = 1
metricsTestLiver.calc_metrics(gtSegs[:,:,:,1], liverPrediction)
lesionPrediction = np.zeros((segPrediction.shape), dtype=np.float32)
lesionPrediction[segPrediction == 2] = 1
metricsTestLesion.calc_metrics(gtSegs[:,:,:,2], lesionPrediction)
if idxTest in randValidBatchIdx:
if args.batch_size > 1:
randValidFrameIdx = np.random.randint(0,args.batch_size-1,1)[0]
else:
randValidFrameIdx = 0
with timer_dict['summary']:
validation_imgs_samples.append((np.copy(images[randValidFrameIdx,:,:,0]), np.copy(segPrediction[randValidFrameIdx,:,:]), np.copy(gtSegs[randValidFrameIdx,:,:,:])))
#timerStats()
#-------------------------------------------------------------------
# Write summaries
#-------------------------------------------------------------------
validation_loss.push(iteration)
summary = sess.run(merged_summary, feed_dict={ucrNet.image: images,
ucrNet.gtSeg: gtSegs})
summary_writer.add_summary(summary, iteration)
metricsRes = metricsTestLiver.summarize_metrics()
validation_metric_liv.add([metricsRes['dice'], metricsRes['precision'], metricsRes['recall'], metricsRes['accuracy']])
validation_metric_liv.push(iteration)
metricsRes = metricsTestLesion.summarize_metrics()
validation_metric_les.add([metricsRes['dice'], metricsRes['precision'], metricsRes['recall'], metricsRes['accuracy']])
validation_metric_les.push(iteration)
validation_imgs.push(iteration, validation_imgs_samples)
validation_imgs_samples = []
summary_writer.flush()
metricsTestLiver.__init__()
metricsTestLesion.__init__()
#-------------------------------------------------------------------
# Save a checktpoint
#-------------------------------------------------------------------
checkpoint = '{}/{}_{}.ckpt'.format(snaps_path, e+1, idx)
model_saver.save(sess, checkpoint)
#print('[i] Checkpoint saved: ' + checkpoint)
start_idx = 0
#-------------------------------------------------------------------
# Save final checktpoint
#-------------------------------------------------------------------
timerStats()
checkpoint = '{}/weights_final.ckpt'.format(snaps_path)
model_saver.save(sess, checkpoint)
print('[i] Checkpoint saved:' + checkpoint)
print('[i] Done training...')
return 0
def predict(test_data_dir, weights_path, out_seg_dir):
#---------------------------------------------------------------------------
# Parse the commandline
#---------------------------------------------------------------------------
args = config
#---------------------------------------------------------------------------
# Find an existing checkpoint
#---------------------------------------------------------------------------
checkpoint_file = weights_path + 'weights_final.ckpt'
if checkpoint_file is None:
print('[!] No checkpoints found, cannot continue!')
return 1
#---------------------------------------------------------------------------
# Configure the training data
#---------------------------------------------------------------------------
print('[i] Configuring the test data...')
if args.testData:
testImagesDir = test_data_dir + '/'
testImagesList = [testImagesDir + f for f in os.listdir(testImagesDir) if os.path.isfile(testImagesDir + f)]
testImagesList.sort()
n_valid_batches = len(testImagesList)
else:
try:
dp = DataPrep(test_data_dir, test_data_dir)
n_valid_batches = int(math.ceil(dp.num_valid/args.batch_size))
print('[i] # test samples: {}'.format(dp.num_valid))
print('[i] # batch size train: {}'.format(args.batch_size))
except (AttributeError, RuntimeError) as e:
print('[!] Unable to load data: ' + str(e))
return 1
#---------------------------------------------------------------------------
# Create the network
#---------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Uploading the model...')
ucrNet = net.UnetCrfRnn(args.batch_size, args.weight_decay, args.bias_decay)
init = tf.global_variables_initializer()
init_local = tf.local_variables_initializer()
sess.run(init)
sess.run(init_local)
if not checkpoint_file is None:
try:
initialize_variables_from_ckpt(sess, checkpoint_file)
#saver = tf.train.Saver()
#saver.restore(sess, checkpoint_file)
except Exception as E:
print E
#-----------------------------------------------------------------------
# Create various helpers
#-----------------------------------------------------------------------
if not os.path.exists(args.logdir):
os.mkdir(args.logdir)
if os.path.exists(os.path.join(args.logdir , config.name)):
shutil.rmtree(os.path.join(args.logdir , config.name))
os.mkdir(os.path.join(args.logdir , config.name))
else:
os.mkdir(os.path.join(args.logdir , config.name))
if not os.path.exists(out_seg_dir):
os.mkdir(out_seg_dir)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(os.path.join(args.logdir , config.name), sess.graph)
#-----------------------------------------------------------------------
# Summaries
#-----------------------------------------------------------------------
validation_metric_liv = MetricsSummary(sess, summary_writer, ['diceLiv', 'precisionLiv', 'recallLiv', 'accuracyLiv'], 1)
validation_metric_les = MetricsSummary(sess, summary_writer, ['diceLes', 'precisionLes', 'recallLes', 'accuracyLes'], 1)
validation_imgs = ImageSummary(sess, summary_writer, 'validation')
validation_loss = LossSummary(sess, summary_writer, 'validation', n_valid_batches)
#---------------------------------------------------------------------------
# metrics initilazing
#---------------------------------------------------------------------------
#metricsTrainLiver = lp_metrics()
#metricsTrainLesion = lp_metrics()
metricsTestLiver = lp_metrics()
metricsTestLesion = lp_metrics()
#-------------------------------------------------------------------
# Validate
#-------------------------------------------------------------------
if args.testData:
description = '[i] Test '
for idxTest, image in enumerate(tqdm(testImagesList, total=n_valid_batches, desc=description, unit='batches', leave=False)):
testImg = cv2.imread(image)
testImg = testImg[:,:,0]
testImg = testImg[np.newaxis,:,:,np.newaxis].astype(np.float32)
[segPrediction] = sess.run([ucrNet.segPrediction], feed_dict={ucrNet.image: testImg})
#-------------------------------------------------------------------
# Dump segmented images
#-------------------------------------------------------------------
segToSave = np.zeros((segPrediction.shape[1], segPrediction.shape[2]), dtype=np.uint8)
segToSave[segPrediction[0] == 1] = 127
segToSave[segPrediction[0] == 2] = 255
testImageName = re.search("{}\/ct(_.+)".format(test_data_dir), image).group(1)
cv2.imwrite('{}/seg{}.png'.format(out_seg_dir, testImageName), segToSave)
print('[i] Done testing...')
else:
iteration = 0
validation_imgs_samples = []
randValidBatchIdx = random.sample(range(n_valid_batches-1), 10)
generator = dp.valid_generator(args.batch_size, args.num_workers)
description = '[i] Valid '
for idxTest, (images, gtSegs) in enumerate(tqdm(generator, total=n_valid_batches, desc=description, unit='batches', leave=False)):
[loss, segPrediction] = sess.run([ucrNet.loss, ucrNet.segPrediction], feed_dict={ucrNet.image: images, ucrNet.gtSeg: gtSegs})
validation_loss.add(loss)
liverPrediction = np.zeros((segPrediction.shape), dtype=np.float32)
liverPrediction[segPrediction == 1] = 1
metricsTestLiver.calc_metrics(gtSegs[:,:,:,1], liverPrediction)
lesionPrediction = np.zeros((segPrediction.shape), dtype=np.float32)
lesionPrediction[segPrediction == 2] = 1
metricsTestLesion.calc_metrics(gtSegs[:,:,:,2], lesionPrediction)
if idxTest in randValidBatchIdx:
if args.batch_size > 1:
randValidFrameIdx = np.random.randint(0,args.batch_size-1,1)[0]
else:
randValidFrameIdx = 0
with timer_dict['summary']:
validation_imgs_samples.append((np.copy(images[randValidFrameIdx,:,:,0]), np.copy(segPrediction[randValidFrameIdx,:,:]), np.copy(gtSegs[randValidFrameIdx,:,:,:])))
#timerStats()
#-------------------------------------------------------------------
# Write summaries
#-------------------------------------------------------------------
validation_loss.push(iteration)
summary = sess.run(merged_summary, feed_dict={ucrNet.image: images,
ucrNet.gtSeg: gtSegs})
summary_writer.add_summary(summary, iteration)
metricsRes = metricsTestLiver.summarize_metrics()
validation_metric_liv.add([metricsRes['dice'], metricsRes['precision'], metricsRes['recall'], metricsRes['accuracy']])
validation_metric_liv.push(iteration)
print "Liver segmentation results: {}".format(metricsRes)
metricsRes = metricsTestLesion.summarize_metrics()
validation_metric_les.add([metricsRes['dice'], metricsRes['precision'], metricsRes['recall'], metricsRes['accuracy']])
validation_metric_les.push(iteration)
print "Lesion segmentation results: {}".format(metricsRes)
validation_imgs.push(iteration, validation_imgs_samples)
validation_imgs_samples = []
summary_writer.flush()
print('[i] Done validating...')
return 0
if __name__ == '__main__':
key = int(raw_input("Please type: train only - 1, both train and predict - 2, predict only - 3\n"))
if key < 3:
train_data_dir = config.train_data_dir
val_data_dir = config.val_data_dir
out_weight_dir = config.snapdir
tf.reset_default_graph()
status = train(train_data_dir, val_data_dir, out_weight_dir)
if status == 1:
sys.exit(status)
if key > 1:
test_data_dir = config.val_data_dir
if key == 2:
weights_path = config.snapdir + '/' + config.name
else:
weights_path = config.weight_path_test
out_seg_dir = config.out_seg_dir
tf.reset_default_graph()
status = predict(test_data_dir, weights_path, out_seg_dir)
if status == 1:
sys.exit(status)