model.py

import tensorflow as tf
from tensorflow.python.keras.engine import training


def create_dnn(seed=42, input_shape=(2381), mc=False):
    """
    This function compiles and returns a Keras model.
    """

    initializer = tf.keras.initializers.GlorotNormal(seed=seed)

    # model = tf.keras.models.Sequential()
    inputs = tf.keras.Input(shape=input_shape)
    x = tf.keras.layers.Dense(2381, activation='elu', kernel_initializer=initializer)(inputs)
    x = tf.keras.layers.LayerNormalization() (x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(1024, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization() (x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(512, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization() (x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(128, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization() (x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    outputs = tf.keras.layers.Dense(1, activation="sigmoid")(x)
    model = tf.keras.Model(inputs, outputs)

    bce = tf.keras.losses.BinaryCrossentropy(from_logits=False)
    # optim = tf.keras.optimizers.RMSprop(lr=1e-3, momentum=0.9)
    optim = tf.keras.optimizers.Adam()

    model.compile(optimizer=optim, loss=bce, 
        metrics=['accuracy'])
#     metrics=['accuracy',
#             tf.keras.metrics.SensitivityAtSpecificity(0.99, name="TPR_01")])

    
    return model


def create_dnn2(seed=42, mc=False):
    """
    This function compiles and returns a Keras model.
    """

    initializer = tf.keras.initializers.GlorotNormal(seed=seed)

    # model = tf.keras.models.Sequential()
    input1 = tf.keras.Input(shape=(2381, ))
    input2 = tf.keras.Input(shape=(1, ))

    x = tf.keras.layers.concatenate([input1, input2])
    x = tf.keras.layers.Dense(2382, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization()(x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(1024, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization()(x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(512, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization()(x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    x = tf.keras.layers.Dense(128, activation='elu', kernel_initializer=initializer)(x)
    x = tf.keras.layers.LayerNormalization()(x)
    x = tf.keras.layers.Dropout(0.3)(x, training=mc)

    outputs = tf.keras.layers.Dense(1)(x)
    out = tf.add(outputs, input2)

    # out = tf.clip_by_value(out, 0, 1)
    out = tf.sigmoid(out)
    model = tf.keras.Model((input1, input2), out)

    bce = tf.keras.losses.BinaryCrossentropy(from_logits=False)
    # mae = tf.keras.losses.MeanAbsoluteError()
    # mse = tf.keras.losses.MeanSquaredError()
    optim = tf.keras.optimizers.Adam()

    model.compile(optimizer=optim, loss=bce, 
        metrics=['accuracy'])
#     metrics=['accuracy',
#             tf.keras.metrics.SensitivityAtSpecificity(0.99, name="TPR_01")])
    return model