create_autoencoder.py

import tensorflow as tf 
from tensorflow.keras.layers import InputLayer, Dense, Input, Dropout,\
BatchNormalization, Flatten
from tensorflow.keras import regularizers
from tensorflow.keras.models import Sequential, Model, load_model
from tensorflow.keras.optimizers import Adam
from functools import partial

def create_ae_model(num_hidden_layers=3, hidden_layer_neurons=64,
                    lr=0.001, kernel_init='lecun_normal', optimizer='adam',
                    noise_at_input=0.05, noise_in_hidden_layer=0.05,
                    activity_regularizer=1e-05,
                    act='relu', dropout=0, input_data_dim=NO_OF_FEATURES):
    
    # assuming all hidden layers have same dimension. 
    dims = [hidden_layer_neurons, hidden_layer_neurons,
            hidden_layer_neurons, hidden_layer_neurons][:num_hidden_layers]

    RegularizedDense = partial(tf.keras.layers.Dense,
                               activation=act,
                               kernel_initializer=kernel_init,
                               activity_regularizer=regularizers.l1(activity_regularizer)
                               )
    
    # We initialze a sequential model.
    autoencoder = Sequential()
    # Add inputs to the model. Notice input_data_dim deafult value is NO_OF_FEATURES.
    autoencoder.add(Input(shape=input_data_dim))
    
    # This is encode phase.
    for i, dim in enumerate(dims):
        autoencoder.add(RegularizedDense(dim))
        autoencoder.add(Dropout(dropout)) if dropout > 0 else ()

    dims.reverse()
    dims = dims[1:]
    
    # This is the decode phase.
    for i, dim in enumerate(dims):
        autoencoder.add(RegularizedDense(dim))
        autoencoder.add(Dropout(dropout)) if dropout > 0 else ()

    autoencoder.add(Dense(input_data_dim, activation='sigmoid'))

    autoencoder.summary()
    
    # Using binary cross entropy loss. It's very common for binary classification. 
    # We will use mean absolute error and mean squared error as additional metrics. 
    autoencoder.compile(optimizer=Adam(learning_rate=lr),
                        loss=tf.keras.losses.binary_crossentropy,
                        metrics=['mae', 'mse'])

    return autoencoder