Conditional Variational Autoencoders with Learnable Conditional Embeddings

This article discusses conditional variational autoencoders (CVAE) and introduces the concept of learnable conditional embeddings in CVAE models. It demonstrates how to implement and train a CVAE model with one-hot encoded conditions and compares it to a model with learnable embeddings. The article also explores the ability of the model to generate images of unseen digits and visualizes the latent space. The use of learnable embeddings is shown to be a useful alternative to one-hot encodings in CVAE models.