medium_tf_model_builder.py

def model_builder(hp):
    if TASK == "r":
        loss_fn = "mean_absolute_error"
    elif TASK == "c":
        if OUTPUT_NODES == 1:
            loss_fn = tf.keras.losses.BinaryCrossentropy(from_logits=True)
        else:
            loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

    if TASK == "r":
        metrics = None
    elif TASK == "c":
        metrics = ["accuracy"]

    kernel_hp = hp.Choice(
        "kernel_regularization", values=[0.01, 0.001, 0.0001, 0.00001]
    )
    activation_hp = hp.Choice("activation", values=["elu", "relu"])
    lr_schedule = tf.keras.optimizers.schedules.InverseTimeDecay(
        0.01, decay_steps=train_size * 1000, decay_rate=1, staircase=False
    )

    body = tf.keras.Sequential()
    current_nodes = num_preprocessed_outputs

    while current_nodes > OUTPUT_NODES:
        body.add(
            tf.keras.layers.Dense(
                current_nodes,
                kernel_regularizer=tf.keras.regularizers.l2(kernel_hp),
                activation=activation_hp,
            )
        )
        body.add(tf.keras.layers.Dropout(0.1))

        current_nodes = current_nodes // 2

    body.add(tf.keras.layers.Dense(OUTPUT_NODES))

    result = body(preprocessed_outputs)

    model = tf.keras.Model(model_inputs, result)

    model.compile(
        loss=loss_fn,
        optimizer=tf.keras.optimizers.Adam(learning_rate=lr_schedule),
        metrics=metrics,
    )

    return model