[P] Wine synthesis using VAE [P]

I have created a VAE model using PyTorch on White Wine dataset.

Basically, the main goal is to discover a brand-new white wine recipe.

It puts all the wines into a latent space, finds the best part where higher bands are located, and then it makes 100 steps with a step size of 0.5, multiplying by 0.96 with every step till it reaches the highest possible score.

The regression model then evaluates the newly generated wine recipe, after first passing the latent coordinates through a decoder, and normalizing them.

I get a score in range of 7.30-7.58.

https://preview.redd.it/bylmwhu92iph1.png?width=249&format=png&auto=webp&s=b97f36f3deca86da85dddbba9b65043978ce4fd3

Here is the visualization of the latent space of all the wines:

https://preview.redd.it/pqjl9pck2iph1.png?width=811&format=png&auto=webp&s=e1128bd7f4f38ef7bd584fd8b8622d0260c90dc1

A small question about the training loop:

https://preview.redd.it/03k4k5yq2iph1.png?width=262&format=png&auto=webp&s=7e0df68447e5ab26d680af08f5b9361b14af168c

Is the loss too large? I know that it never could reach perfect zero by how do I know if the loss is good enough? After reaching the plato? I use MSELoss.

The repo itself: https://github.com/theaidenmax/tabular-vae-wine-generator

This is my second project in VAE (after MNIST), and I really liked idea of autoencoders.

And I also would be very grateful for your tips, etc, since this is one of my first steps in ML/DL.