def train_top_model(self, train_data_dir, val_data_dir, epochs=defaults['epochs'],
batch_size=defaults['batch_size']):
train_generator = util.get_generator(train_data_dir, self.img_width, self.img_height, batch_size)
val_generator = util.get_generator(val_data_dir, self.img_width, self.img_height, batch_size)
train_data = util.load_bottleneck_features(self.train_bottleneck_features_path)
val_data = util.load_bottleneck_features(self.val_bottleneck_features_path)
num_classes = train_generator.num_classes
train_labels = train_generator.classes
train_labels = to_categorical(train_labels, num_classes=num_classes)
val_labels = val_generator.classes
val_labels = to_categorical(val_labels, num_classes=num_classes)
model = self.get_top_model(train_data.shape[1:], num_classes)
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
util.save_model_plot(self.top_model_plot_path, model)
history = model.fit(train_data, train_labels,
epochs=epochs,
batch_size=batch_size,
validation_data=(val_data, val_labels))
model.save(self.model_path)
util.save_history(self.history_path, history)
util.eval_model_loss_acc(model, val_data, val_labels, batch_size)
def fine_tune(self, train_data_dir, val_data_dir, epochs, batch_size):
train_generator = util.get_categorical_generator(
train_data_dir, self.img_width, self.img_height, batch_size)
val_generator = util.get_categorical_generator(val_data_dir,
self.img_width,
self.img_height,
batch_size)
base_model = self.get_base_model()
base_model.trainable = True
set_trainable = False
for layer in base_model.layers:
if layer.name == 'block5_conv1':
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
top_model = self.get_top_model(base_model.output_shape[1:],
train_generator.num_classes)
top_model.load_weights(self.top_model_weights_path)
top_model.trainable = False
model = Sequential()
model.add(base_model)
model.add(top_model)
model.compile(optimizer=SGD(lr=1e-4, momentum=0.9),
loss='categorical_crossentropy',
metrics=['accuracy'])
util.save_model_plot(self.model_plot_path, model)
history = model.fit_generator(
train_generator,
verbose=1,
steps_per_epoch=train_generator.samples / batch_size,
epochs=epochs,
validation_data=val_generator,
validation_steps=val_generator.samples / batch_size)
model.save_weights(self.model_weights_path)
model.save(self.model_path)
util.save_history(self.history_path, history)