Hands-On Neural Networks with Keras

Training the autoencoder

Next, we simply fit our autoencoder network, just as we've done with other networks countless times before. We chose this model to be trained for 50 epochs, in batches of 256 images, before weight updates to our network nodes are performed. We also shuffle our data during training. As we already know, doing so ensures some variance reduction among batches, thereby improving the generalizability for our model:

Finally, we also defined the validation data using our test set, just to be able to compare how well our model does on unseen examples, at the end of each epoch. Do remember that in normal machine learning workflows, ...