Training the denoising network

We will use the same convolutional autoencoder architecture for this task. However, we will reinitialize the model and train it from scratch once again, this time with the noisy input variables:

autoencoder.compile(optimizer='adam', loss='mse')
autoencoder.fit(x_train_noisy, x_train, epochs=50, batch_size=20,
                shuffle=True, verbose=1)

Epoch 1/50875/875 [==============================] - 7s 8ms/step - loss: 0.0449
Epoch 2/50
875/875 [==============================] - 6s 7ms/step - loss: 0.0212
Epoch 3/50
875/875 [==============================] - 6s 7ms/step - loss: 0.0185
Epoch 4/50
875/875 [==============================] - 6s 7ms/step - loss: 0.0169

As we can see, the loss converges much more reluctantly in the case of the denoising autoencoder than for our previous experiments. This is naturally the case, as...