About Deep Tabular Augmentation for Regression Tasks

[Yesim7]

Hello Lasse (@lschmiddey ),

I am Yesim. I read your blog and wanna ask something about "Deep Tabular Augmentation for Regression Tasks". I didn't understand your some steps. For example, DataBunch, Autoencoder. Are they running with random parameters? Or how did you determine the value you gave to VAE_arch? Could you help me? Could you suggest a resource that has been explained in detail?

Sincerely,

[lschmiddey]

Hi Yesim,

I suggest (if you havent already) reading the last of my blogposts on Deep Tabular Augmentation: https://lschmiddey.github.io/fastpages_/2022/04/23/DataAugmentation_for_Regression_Tasks.ipynb.html

But I will also try to explain it here. The DataBunch class is basically just a helper class for bundling pytorch datasets and pytorch dataloaders. You could also provide your own dataloader to the (model)class Autoencoder. The only thing it needs to be is a pytorch dataloader, otherwise the training will not work. How the DataBunch is defined can be found here: https://github.com/lschmiddey/deep_tabular_augmentation/blob/main/deep_tabular_augmentation/dataloaders.py -> I provided the DataBunch with reasonable default values (like batch_size (bs) of 128) but you can easily change this:

data = dta.DataBunch(*dta.create_loaders(datasets, bs=1024))

When it comes to the model the architecture which I usually use is:(VAE_arch = [50, 12, 12]). However, this depends on the data you have and you can play around with it. Usually, with more variables you can also try using higher values here, but I found this architecture being a good default architecture.

When it comes to understanding VAE a bit better, I suggest this article: https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73

Usually VAE is used for images (for example VAE is part of the stable diffusion model) but the general idea can also be applied to tabular data.

Cheers, Lasse

[Yesim7]

Hello again Lasse (@lschmiddey) ,

Thank you for the links you shared. There are a few things that still come to my mind. I have a physics experiment dataset with two columns as independent variables and one dependent variable. The first of the independent variables consists of positive integers between 0 and 10. The second of the independent variables consists of numbers floating between 0 and 1. I also have the variable y, which is dependent on them. For example, x1, x2, y 1, 0.02, 150.47 1, 0.3, 147.60 1, 0.4, 140.5 ... 10, 0.01, 162.8 10, 0.4, 145.3 10, 0.5, 178.3 I have a dataset as above.

I want to get the y value for all values in between using the deep_tabular_augmentation module you shared. I expected values to be produced between 1 and 10 for the first column and between 0.02 and 0.5 for the second column and positive. However, when I import the module and run the code, I get negative and random values. Actually I want to do something like interpolation. I hope I was able to explain. I would be grateful if you could help me with this.

Sincerely,

[lschmiddey]

I will try to reproduce it and inform you about my findings :)

[Yesim7]

Thank you so much, waiting for your return:)

[Yesim7]

Hi @lschmiddey , Did you have time to try?:)