The discovery of this idea in the original 2013 research paper ("Auto-Encoding Variational Bayes" by D.P. Kingma and M. Welling) was the key to enabling VAEs in practice. Training a Variational ...

The loss function first computes binary cross entropy loss between the source x and the reconstructed x and stores that single tensor value as bce. ... Listing 3: Training a Variational Autoencoder.