_fcn.py

__author__ = "James Large, Withington"

from tensorflow import keras

from sktime_dl.regression._regressor import BaseDeepRegressor
from sktime_dl.networks._fcn import FCNNetwork
from sktime_dl.utils import check_and_clean_data, \
    check_and_clean_validation_data


class FCNRegressor(BaseDeepRegressor, FCNNetwork):
    """Fully convolutional neural network (FCN).

    Adapted from the implementation from Fawaz et. al

    https://github.com/hfawaz/dl-4-tsc/blob/master/classifiers/fcn.py

    Network originally defined in:

    @inproceedings{wang2017time, title={Time series classification from
    scratch with deep neural networks: A strong baseline}, author={Wang,
    Zhiguang and Yan, Weizhong and Oates, Tim}, booktitle={2017
    International joint conference on neural networks (IJCNN)}, pages={
    1578--1585}, year={2017}, organization={IEEE} }

    :param nb_epochs: int, the number of epochs to train the model
    :param batch_size: int, specifying the length of the 1D convolution
     window
    :param callbacks: list of tf.keras.callbacks.Callback objects
    :param random_state: int, seed to any needed random actions
    :param verbose: boolean, whether to output extra information
    :param model_name: string, the name of this model for printing and
     file writing purposes
    :param model_save_directory: string, if not None; location to save
     the trained keras model in hdf5 format
    """

    def __init__(
            self,
            nb_epochs=2000,
            batch_size=16,
            callbacks=None,
            random_state=0,
            verbose=False,
            model_name="fcn_regressor",
            model_save_directory=None,
    ):
        super(FCNRegressor, self).__init__(
            model_name=model_name, model_save_directory=model_save_directory
        )

        self.verbose = verbose
        self._is_fitted = False

        # calced in fit
        self.input_shape = None
        self.history = None

        # predefined
        self.nb_epochs = nb_epochs
        self.batch_size = batch_size

        self.callbacks = callbacks
        self.random_state = random_state
        self.verbose = verbose
        self._is_fitted = False

    def build_model(self, input_shape, **kwargs):
        """
        Construct a compiled, un-trained, keras model that is ready for
         training
        ----------
        input_shape : tuple
            The shape of the data fed into the input layer
        Returns
        -------
        output : a compiled Keras Model
        """
        input_layer, output_layer = self.build_network(input_shape, **kwargs)

        output_layer = keras.layers.Dense(units=1)(output_layer)

        model = keras.models.Model(inputs=input_layer, outputs=output_layer)

        model.compile(
            loss="mean_squared_error",
            optimizer=keras.optimizers.Adam(),
            metrics=["mean_squared_error"],
        )

        # if user hasn't provided a custom ReduceLROnPlateau via init
        # already, add the default from literature
        if self.callbacks is None:
            self.callbacks = []

        if not any(
                isinstance(callback, keras.callbacks.ReduceLROnPlateau)
                for callback in self.callbacks
        ):
            reduce_lr = keras.callbacks.ReduceLROnPlateau(
                monitor="loss", factor=0.5, patience=50, min_lr=0.0001
            )
            self.callbacks.append(reduce_lr)

        return model

    def fit(self, X, y, input_checks=True, validation_X=None,
            validation_y=None, **kwargs):
        """
        Fit the regressor on the training set (X, y)
        ----------
        X : a nested pd.Dataframe, or (if input_checks=False) array-like of
        shape = (n_instances, series_length, n_dimensions)
            The training input samples. If a 2D array-like is passed,
            n_dimensions is assumed to be 1.
        y : array-like, shape = [n_instances]
            The training data class labels.
        input_checks : boolean
            whether to check the X and y parameters
        validation_X : a nested pd.Dataframe, or array-like of shape =
        (n_instances, series_length, n_dimensions)
            The validation samples. If a 2D array-like is passed,
            n_dimensions is assumed to be 1.
            Unless strictly defined by the user via callbacks (such as
            EarlyStopping), the presence or state of the validation
            data does not alter training in any way. Predictions at each epoch
            are stored in the model's fit history.
        validation_y : array-like, shape = [n_instances]
            The validation class labels.
        Returns
        -------
        self : object
        """
        X = check_and_clean_data(X, y, input_checks=input_checks)

        validation_data = \
            check_and_clean_validation_data(validation_X, validation_y)

        # ignore the number of instances, X.shape[0],
        # just want the shape of each instance
        self.input_shape = X.shape[1:]

        self.batch_size = int(max(1, min(X.shape[0] / 10, self.batch_size)))

        self.model = self.build_model(self.input_shape)

        if self.verbose:
            self.model.summary()

        self.history = self.model.fit(
            X,
            y,
            batch_size=self.batch_size,
            epochs=self.nb_epochs,
            verbose=self.verbose,
            callbacks=self.callbacks,
            validation_data=validation_data,
        )

        self.save_trained_model()
        self._is_fitted = True

        return self