whatashot / excelformer Goto Github PK

Hi, I was studying the provided code in run_default_config_excel.py, and I realized that the categorical features are encoded using CatBoostEncoder. Then, if you apply this encoding process, you set that X_cat = 0, because they are transformed to numerical features. So I have these questions:

1. If I dont use this encoding process, ExcelFormer architecture will deal with these categorical features?
2. 'cardinalities' refers to the number of categorical features that I have in my dataset?
3. Do I need to set WHICH are the categorical features that I have, if I dont apply CatBoostEncoder?

I appreciate your attention and you've done a great work!

I am so glad to find your work, and finally see a nn that beats tree models.

Would it possible for you to upload the "/data2/yanjianhuan/research/data/tabular/data" to cloud? So it would be easier for me to run the code. I would like to run your code on your data first to do a sanity check first before trying on my data.

Hi<
Congrats on your great research paper,
I just ran the code for CA and in the best scenario, I got ~-0.445 which is far from -0.433 reported in the draft! Have you considered any specific constraints in your model during training? I used the default params recommended in the repo.
BWs,

Intrigued by reading the paper. Is there any code available?

Hi, great work!

Is there any chance for ExcelFormer to support scikit-learn fit API, so we can use something like ExcelFormerClassifier.fit().

In your another repo, TabCaps, model can be fitted by TabCapsClassifier.fit(), which is good.

It can help compare performance of multiple tabular models if they are all in one format, thanks!