model.py

import tensorflow as tf
from attention_layer import AttentionLayer


def build_model(window_size, filters, kernel_size, units):
    input_data = tf.keras.Input(shape=(window_size, 1))

    # CLASSIFICATION SUBNETWORK
    x = tf.keras.layers.Conv1D(filters=30, kernel_size=10, activation='relu')(input_data)
    x = tf.keras.layers.Conv1D(filters=30, kernel_size=8, activation='relu')(x)
    x = tf.keras.layers.Conv1D(filters=40, kernel_size=6, activation='relu')(x)
    x = tf.keras.layers.Conv1D(filters=50, kernel_size=5, activation='relu')(x)
    x = tf.keras.layers.Conv1D(filters=50, kernel_size=5, activation='relu')(x)
    x = tf.keras.layers.Conv1D(filters=50, kernel_size=5, activation='relu')(x)
    x = tf.keras.layers.Flatten()(x)
    x = tf.keras.layers.Dense(units=1024, activation='relu', kernel_initializer='he_normal')(x)
    classification_output = tf.keras.layers.Dense(units=window_size, activation='sigmoid', name="classification_output")(x)

    #REGRESSION SUBNETWORK
    y = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel_size, activation='relu')(input_data)
    y = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel_size, activation='relu')(y)
    y = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel_size, activation='relu')(y)
    y = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel_size, activation='relu')(y)
    y = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(units, activation="tanh", return_sequences=True), merge_mode="concat")(y)
    y, weights = AttentionLayer(units=units)(y)
    y = tf.keras.layers.Dense(units, activation='relu')(y)
    regression_output = tf.keras.layers.Dense(window_size, activation='relu', name="regression_output")(y)

    output = tf.keras.layers.Multiply(name="output")([regression_output, classification_output])

    full_model = tf.keras.Model(inputs=input_data, outputs=[output, classification_output], name="LDwA")
    attention_model = tf.keras.Model(inputs=input_data, outputs=weights)

    optimizer = tf.keras.optimizers.SGD(learning_rate=0.01, momentum=0.9)

    full_model.compile(optimizer=optimizer, loss={
        "output": tf.keras.losses.MeanSquaredError(),
        "classification_output": tf.keras.losses.BinaryCrossentropy()})

    return full_model, attention_model