Compiling and Training the Autoencoder

compile_and_train.py

def lr_scheduler1(epoch, lr):
    return lr * (0.995 ** epoch)

training_generator = model.DAEGenerator(features=X_train_all_standard,
                                      batch_size=256,
                                      swap_noise_rate=0.15)

autoencoder.compile(loss='mean_squared_error', optimizer=optimizers.Adam(lr=0.0005))

es = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=20, min_delta=0.001)
lrs = LearningRateScheduler(lr_scheduler1, verbose=1)

cp = ModelCheckpoint(filepath='autoencoder.h5,
                     monitor='val_loss',
                     save_best_only=True,
                     verbose=0)

history = autoencoder.fit_generator(generator=training_generator, 
                                    validation_data=(X_val_standard, X_val_standard),
                                    use_multiprocessing=False, max_queue_size=500,
                                    callbacks=[es, lrs, cp], epochs=500)