Read This Before Building Custom Keras Layers & Models

This article is a practical tour of Keras internals. It shows how to write custom Layers (with trainable and non-trainable weights), lazily create weights in build(), compose layers into reusable blocks, and register auxiliary losses via add_loss(). It covers serialization (get_config/from_config), privileged training and mask args, and when to use Model vs Layer. It ends with an end-to-end Variational Autoencoder on MNIST—implemented as nested, serializable layers, trained via both a custom loop and model.fit().