I’ve been playing around with autoencoders in Pytorch lightning mainly to figure out how they are setup and running them.
I’ve seen @muellerzr has some autoencoders working with tabular data (keen to explore) but are there any other uses of autoencoders or generative models in fastai?
I’m curious as I want to explore generating synthetic data.
I worked with @muellerzr on the first tabular autoencoder, but this is not really a generative model. I also did a Variational Autoencoder version for tabular data here if you are interested https://github.com/EtienneT/TabularVAE/blob/master/TabularVAE.ipynb. Basically once your model is fitted, you can sample a vector from the standard normal distribution, pass it to the decoder of the network and it should generate new unseen tabular data. The generated data should make sense in terms of data distribution etc…