def train(fold, train_patient_indexes, val_patient_indexes):
log_dir = 'fold_' + str(fold) + '/'
if not os.path.isdir(log_dir):
os.mkdir(log_dir)
num_slices_train = len(train_patient_indexes) * 189
num_slices_val = len(val_patient_indexes) * 189
# Create model
K.clear_session()
model = create_xception_unet_n(
input_shape=input_shape,
pretrained_weights_file=pretrained_weights_file)
model.compile(optimizer=Adam(lr=1e-3), loss=get_loss, metrics=[dice])
# Get callbacks
checkpoint = ModelCheckpoint(
log_dir + 'ep={epoch:03d}-loss={loss:.3f}-val_loss={val_loss:.3f}.h5',
verbose=1,
monitor='val_loss',
save_weights_only=True,
save_best_only=True,
period=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
min_delta=1e-3,
patience=3,
verbose=1)
early_stopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=5,
verbose=1)
csv_logger = CSVLogger(log_dir + 'record.csv')
tensorboard = TensorBoard(log_dir=log_dir)
# train the model
model.fit_generator(
create_train_date_generator(patient_indexes=train_patient_indexes,
h5_file_path=data_file_path,
batch_size=batch_size),
steps_per_epoch=max(1, num_slices_train // batch_size),
validation_data=create_val_date_generator(
patient_indexes=val_patient_indexes,
h5_file_path=data_file_path,
batch_size=9),
validation_steps=max(1, num_slices_val // 9),
epochs=num_epoch,
initial_epoch=0,
callbacks=[
checkpoint, reduce_lr, early_stopping, tensorboard, csv_logger
])
model.save_weights(log_dir + 'trained_final_weights.h5')
#save model itself
model.save(os.path.join(log_dir, 'trained_final_model'))
# model.save(os.path.join(log_dir, 'trained_final_model.h5'))
# Evaluate model
predicts = []
labels = []
f = create_val_date_generator(patient_indexes=val_patient_indexes,
h5_file_path=data_file_path)
for _ in range(num_slices_val):
img, label = f.__next__()
predicts.append(model.predict(img))
labels.append(label)
predicts = np.array(predicts)
labels = np.array(labels)
score_record = get_score_from_all_slices(labels=labels, predicts=predicts)
# save score
df = pd.DataFrame(score_record)
df.to_csv(os.path.join(log_dir, 'score_record.csv'), index=False)
# print score
mean_score = {}
for key in score_record.keys():
print('In fold ', fold, ', average', key, ' value is: \t ',
np.mean(score_record[key]))
mean_score[key] = np.mean(score_record[key])
# exit training
K.clear_session()
return mean_score
output_path_gt_final = os.path.join(
output_dir, "".join(
["patient_",
str(patient_index), "_gt_",
str(num_slices_val - 1)]))
if (os.path.isfile("".join([output_path_gt_final, ".npy"]))):
print("".join([
"output_path_gt_final (",
str("".join([output_path_gt_final, ".npy"])),
") already exists. Skipping patient ",
str(patient_index)
]),
flush=True)
continue
f = create_val_date_generator(patient_indexes=val_patient_indexes,
h5_file_path=data_file_path)
# img = nib.load(sample_input).get_fdata()
# print("img shape")
# print(np.shape(img))#(197, 233, 189)
# sample_img = img[:,:,90]
print("create_val_date_generator finished", flush=True)
print("".join(["num_slices_val: ", str(num_slices_val)]), flush=True)
for slice_index in np.arange(num_slices_val):
print("".join([
"patient_index | slice_index: ",
str(patient_index), "\t|\t",
str(slice_index)
]),
flush=True)
output_path_seg = os.path.join(
output_dir, "".join(