tabnet
tabnet copied to clipboard
hyperparameter tuning questions
Sorry if my 2 questions below sound noobish:
- In the original TabNet paper under "Guidelines for hyperparameters" section, it talks about
"An optimal choice of γ can have a major role on the performance."
Can someone tell me which one of the hyperparameters in tabnet_config deals with this γ?
- Is there a package that can help auto tune the hyperparameters for a Tabnet?
Hello @3SMMZRjWgS
-
The
gamma
parameter in the paper and original code has been renamedfeature_reusage
intabnet_config()
function. It defaults to 1.3 -
You can tune this hyper-parameter through any method you want using the {workflow} package for example. For his you need to use the
tabnet()
command to fit tabnet models using {parsnip} machinery. You can derive the workflow test if that helps.
@cregouby: Thank you for the explanation and showing the way!