Hey, thanks for your question! I’d suggest passing in check_val_every_n_epoch=1 into train() in order to compute validation losses as well. If you find that elbo_validation, validation_loss, and reconstruction_loss_validation are converging at the end of training, then your model is probably fine. Otherwise, you can tweak the learning rate or number of epochs depending on what’s going on. Hope this helps!
1 Like