options['weight_paths'] = os.path.join(options['train_folder'], scan,
'nets')
options['load_weights'] = False
options['test_scan'] = scan
# train the model for the current scan
print("> CNN: training net with %d subjects" %
(len(list(train_x_data.keys()))))
# --------------------------------------------------
# initialize the CNN and train the classifier
# --------------------------------------------------
model = cascade_model(options)
model = train_cascaded_model(model, train_x_data, train_y_data, options,
CURRENT_PATH)
print("> INFO: training time:", round(time.time() - seg_time), "sec")
print("> INFO: total pipeline time: ", \
round(time.time() - total_time), "sec")
# --------------------------------------------------
# test the current scan
# --------------------------------------------------
test_x_data = {
scan: {
m: os.path.join(options['train_folder'], scan, 'tmp', n)
for m, n in zip(options['modalities'], options['x_names'])
}
}
print("test_x_data : ", test_x_data)
def train_network(options):
"""
Train the CNN network given the options passed as parameter
"""
# set GPU mode from the configuration file. Trying to update
# the backend automatically from here in order to use either theano
# or tensorflow backends
from CNN.base import train_cascaded_model
from CNN.build_model import cascade_model
# define the training backend
define_backend(options)
scan_list = os.listdir(options['train_folder'])
scan_list.sort()
options['task'] = 'training'
options['train_folder'] = os.path.normpath(options['train_folder'])
for scan in scan_list:
total_time = time.time()
# --------------------------------------------------
# move things to a tmp folder before starting
# --------------------------------------------------
options['tmp_scan'] = scan
current_folder = os.path.join(options['train_folder'], scan)
options['tmp_folder'] = os.path.normpath(
os.path.join(current_folder, 'tmp'))
# preprocess scan
preprocess_scan(current_folder, options)
# --------------------------------------------------
# WM MS lesion training
# - configure net and train
# --------------------------------------------------
seg_time = time.time()
print "> CNN: Starting training session"
# select training scans
train_x_data = {
f: {
m: os.path.join(options['train_folder'], f, 'tmp', n)
for m, n in zip(options['modalities'], options['x_names'])
}
for f in scan_list
}
train_y_data = {
f: os.path.join(options['train_folder'], f, 'tmp', 'lesion.nii.gz')
for f in scan_list
}
options['weight_paths'] = os.path.join(CURRENT_PATH, 'nets')
options['load_weights'] = False
# train the model for the current scan
print "> CNN: training net with %d subjects" % (len(train_x_data.keys()))
# --------------------------------------------------
# initialize the CNN and train the classifier
# --------------------------------------------------
model = cascade_model(options)
model = train_cascaded_model(model, train_x_data, train_y_data, options)
print "> INFO: training time:", round(time.time() - seg_time), "sec"
print "> INFO: total pipeline time: ", round(time.time() -
total_time), "sec"
print "> INFO: All processes have been finished. Have a good day!"
if not os.path.exists(os.path.join(options['train_folder'], scan, 'nets')):
os.mkdir(os.path.join(options['train_folder'], scan, 'nets'))
options['weight_paths'] = os.path.join(options['train_folder'], scan,
'nets')
options['load_weights'] = False
options['test_scan'] = scan
# train the model for the current scan
print "> CNN: training net with %d subjects" % (len(train_x_data.keys()))
# --------------------------------------------------
# initialize the CNN and train the classifier
# --------------------------------------------------
model = cascade_model(options)
model = train_cascaded_model(model, train_x_data, train_y_data, options)
print "> INFO: training time:", round(time.time() - seg_time), "sec"
print "> INFO: total pipeline time: ", round(time.time() -
total_time), "sec"
# --------------------------------------------------
# test the current scan
# --------------------------------------------------
test_x_data = {
scan: {
m: os.path.join(options['train_folder'], scan, 'tmp', n)
for m, n in zip(options['modalities'], options['x_names'])
}
}
def train_network(options):
"""
Train the CNN network given the options passed as parameter
"""
# set GPU mode from the configuration file.
# So, this has to be updated before calling
# the CNN libraries if the default config "~/.theanorc" has to be replaced.
if options['mode'].find('cuda') == -1 and options['mode'].find(
'gpu') == -1:
os.environ[
'THEANO_FLAGS'] = 'mode=FAST_RUN,device=cpu,floatX=float32,optimizer=fast_compile'
else:
os.environ['THEANO_FLAGS'] = 'mode=FAST_RUN,device=' + options[
'mode'] + ',floatX=float32,optimizer=fast_compile'
from CNN.base import train_cascaded_model
from CNN.build_model import cascade_model
scan_list = os.listdir(options['train_folder'])
scan_list.sort()
options['task'] = 'training'
options['train_folder'] = os.path.normpath(options['train_folder'])
for scan in scan_list:
total_time = time.time()
# --------------------------------------------------
# move things to a tmp folder before starting
# --------------------------------------------------
options['tmp_scan'] = scan
current_folder = os.path.join(options['train_folder'], scan)
options['tmp_folder'] = os.path.normpath(
os.path.join(current_folder, 'tmp'))
# preprocess scan
preprocess_scan(current_folder, options)
# --------------------------------------------------
# WM MS lesion training
# - configure net and train
# --------------------------------------------------
seg_time = time.time()
print "> CNN: Starting training session"
# select training scans
train_x_data = {
f: {
m: os.path.join(options['train_folder'], f, 'tmp', n)
for m, n in zip(options['modalities'], options['x_names'])
}
for f in scan_list
}
train_y_data = {
f: os.path.join(options['train_folder'], f, 'tmp', 'lesion.nii.gz')
for f in scan_list
}
options['weight_paths'] = os.path.join(CURRENT_PATH, 'nets')
options['load_weights'] = False
# train the model for the current scan
print "> CNN: training net with %d subjects" % (len(train_x_data.keys()))
# --------------------------------------------------
# initialize the CNN and train the classifier
# --------------------------------------------------
model = cascade_model(options)
model = train_cascaded_model(model, train_x_data, train_y_data, options)
print "> INFO: training time:", round(time.time() - seg_time), "sec"
print "> INFO: total pipeline time: ", round(time.time() -
total_time), "sec"
print "> INFO: All processes have been finished. Have a good day!"