# build model
unet = UNet(input_size=(img_width, img_height, 1),
n_filters=64,
pretrained_weights=pretrained_weights)
unet.build()
# creating a callback, hence best weights configurations will be saved
model_checkpoint = ModelCheckpoint(filepath=model_path,
monitor='val_loss',
verbose=1)
#lr schedule
reduceLR = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=7)
#callbacks of CsvLogger
logs_path = os.path.join(SETTINGS_JSON['UNET_LOG_DIR'], "log.csv")
csvlogger = CSVLogger(logs_path)
callbacks = [csvlogger, model_checkpoint, reduceLR]
# model training
# steps per epoch should be equal to number of samples in database divided by batch size
# in this case, it is 528 / 2 = 264
unet.fit_generator(
generator=train_gen,
steps_per_epoch=len(os.listdir(os.path.join(train_path, 'images'))) //
batch_size,
epochs=200,
validation_data=val_gen,
validation_steps=val_samples // batch_size,
callbacks=callbacks)
pretrained_weights = model_weights_name
else:
pretrained_weights = None
# build model
unet = UNet(input_size=(img_width, img_height, 1),
n_filters=64,
pretrained_weights=pretrained_weights)
unet.build()
# creating a callback, hence best weights configurations will be saved
model_checkpoint = unet.checkpoint(model_name)
# model training
# steps per epoch should be equal to number of samples in database divided by batch size
# in this case, it is 528 / 2 = 264
unet.fit_generator(train_gen,
steps_per_epoch=264,
epochs=5,
callbacks=[model_checkpoint])
# saving model weights
unet.save_model(model_weights_name)
# generated testing set
test_gen = test_generator(test_path, 30, img_size)
# display results
results = unet.predict_generator(test_gen, 30, verbose=1)
save_results(save_path, results)
