KerasWrappers

Wrappers around the Sequential and Functional API of Keras
by Alexander Braekevelt.

Importing the wrappers

from keras_wrappers import SequentialWrapper, ModelWrapper

Creating a model

Sequential API

class MyModel(SequentialWrapper):

    def __init__(self):
        model = super().__init__(name='my_model')
	model.add(Dense(32, input_dim=784))
	model.add(Activation('softmax'))
	model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

    @Override
    def preprocess_x(self, data):
	# Optional preprocessing
        return super().preprocess_x(data)

    @Override
    def preprocess_y(self, data):
	# Optional preprocessing
        return super().preprocess_y(data)

    @Override
    def postprocess(self, data):
	# Optional postprocessing
        return super().postprocess(data)


my_model = MyModel()

Functional API

class MyModel(ModelWrapper):

    def __init__(self):
	a = Input(shape=(32,))
	b = Dense(32)(a)
        model = super().__init__(inputs=a, outputs=b, name='my_model')
	model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

    @Override
    def preprocess_x(self, data):
	# Optional preprocessing
        return super().preprocess_x(data)

    @Override
    def preprocess_y(self, data):
	# Optional preprocessing
        return super().preprocess_y(data)

    @Override
    def postprocess(self, data):
	# Optional postprocessing
        return super().postprocess(data)


my_model = MyModel()

Training

Training saves epochs (if not interrupted) and applies both types of preprocessing.

my_model.train(x, y)

from keras.preprocessing.image import ImageDataGenerator
generator = ImageDataGenerator(
            width_shift_range=0.1,
            height_shift_range=0.1,
            horizontal_flip=True
          )
my_model.train_generator(generator, x, y, batch_size=64, epochs=5)

Plotting

Plots loss and accuracy of all epochs combined.

my_model.plot_history(log_y=False)

Predicting

Predicting applies preprocessing and postprocessing.

prediction = my_model.predict_one(single_y)
predictions = my_model.predict_all(multiple_y)

Saving

Saves both the model weights and the history.

my_model.save_model()

Loading

Restores the model weights and the history. (Requires same model architecture.)

my_model.load_model('my_model_35_epochs')

Other methods

All methods of the original model also exist for the wrappers.

braekevelt/KerasWrappers