def fit(self, batch_size=128, nb_epochs=100, save_history=True, history_fn="Model History.txt") -> Model:
"""
Standard method to train any of the models.
"""
samples_per_epoch = img_utils.image_count()
val_count = img_utils.val_image_count()
if self.model == None: self.create_model(batch_size=batch_size)
callback_list = [callbacks.ModelCheckpoint(self.weight_path, monitor='val_PSNRLoss', save_best_only=True,
mode='max', save_weights_only=True, verbose=2)]
if save_history:
callback_list.append(HistoryCheckpoint(history_fn))
if K.backend() == 'tensorflow':
log_dir = './%s_logs/' % self.model_name
tensorboard = TensorBoardBatch(log_dir, batch_size=batch_size)
callback_list.append(tensorboard)
print("Training model : %s" % (self.__class__.__name__))
self.model.fit_generator(img_utils.image_generator(train_path, scale_factor=self.scale_factor,
small_train_images=self.type_true_upscaling,
batch_size=batch_size),
steps_per_epoch=samples_per_epoch // batch_size + 1,
epochs=nb_epochs, callbacks=callback_list,
validation_data=img_utils.image_generator(validation_path,
scale_factor=self.scale_factor,
small_train_images=self.type_true_upscaling,
batch_size=batch_size),
validation_steps=val_count // batch_size + 1)
return self.model
def fit(self, batch_size=128, nb_epochs=100, save_history=True, history_fn="Model History.txt") -> Model:
"""
Standard method to train any of the models.
"""
samples_per_epoch = img_utils.image_count()
val_count = img_utils.val_image_count()
if self.model == None: self.create_model(batch_size=batch_size)
callback_list = [callbacks.ModelCheckpoint(self.weight_path, monitor='val_PSNRLoss', save_best_only=True,
mode='max', save_weights_only=True)]
if save_history: callback_list.append(HistoryCheckpoint(history_fn))
print("Training model : %s" % (self.__class__.__name__))
self.model.fit_generator(img_utils.image_generator(train_path, scale_factor=self.scale_factor,
small_train_images=self.type_true_upscaling,
batch_size=batch_size),
samples_per_epoch=samples_per_epoch,
nb_epoch=nb_epochs, callbacks=callback_list,
validation_data=img_utils.image_generator(validation_path,
scale_factor=self.scale_factor,
small_train_images=self.type_true_upscaling,
batch_size=batch_size),
nb_val_samples=val_count)
return self.model
def fit(self, scale_factor, weight_fn, batch_size=128, nb_epochs=100, small_train_images=False,
save_history=True, history_fn="Model History.txt") -> Model:
"""
Standard method to train any of the models.
"""
samples_per_epoch = img_utils.image_count()
val_count = img_utils.val_image_count()
if self.model == None: self.create_model(batch_size=batch_size)
callback_list = [callbacks.ModelCheckpoint(weight_fn, monitor='val_PSNRLoss', save_best_only=True, mode='max', save_weights_only=True),]
if save_history: callback_list.append(HistoryCheckpoint(history_fn))
print("Training model : %s" % (self.__class__.__name__))
self.model.fit_generator(img_utils.image_generator(train_path, scale_factor=scale_factor,
small_train_images=small_train_images,
batch_size=batch_size),
samples_per_epoch=samples_per_epoch,
nb_epoch=nb_epochs, callbacks=callback_list,
validation_data=img_utils.image_generator(validation_path,
scale_factor=scale_factor,
small_train_images=small_train_images,
batch_size=batch_size),
nb_val_samples=val_count)
return self.model
joint_model.add_loss(l2_weight * teacher_student_loss)
# perceptual loss
with K.name_scope('perceptual_loss'):
perceptual_weight = 2.
perceptual_loss = K.sum(
K.square(
gram_matrix(teacher_output_tensor) -
gram_matrix(student_output_tensor)))
joint_model.add_loss(perceptual_weight * perceptual_loss)
joint_model.compile(optimizer='adam', loss=zero_loss)
# train student model using teacher model
samples_per_epoch = img_utils.image_count()
val_count = img_utils.val_image_count()
weight_path = 'weights/joint_model (%s) %dX.h5' % (teacher_model.model_name,
scale_factor)
history_fn = 'Joint_model_training.txt'
train_path = img_utils.output_path
validation_path = img_utils.validation_output_path
path_X = img_utils.output_path + "X/"
path_Y = img_utils.output_path + "y/"
callback_list = [
ModelCheckpoint(weight_path,
monitor='val_loss',
save_best_only=True,
mode='min',