Add an rmsle evaluation metric?

Feature request

The implementation of a root mean squared logarithmic error regression loss (rmsle) evaluation metric as an out-of-the-box option in TabNet, as it is frequently used in kaggle competitions. I know one can write a little python class, to do this, but having it as an option would be even easier!

What is the expected behavior?

return the rmsle

What is motivation or use case for adding/changing the behavior?

This particular metric is often used in kaggle competitions.

How should this be implemented in your opinion?

A simple addition+modification based on the code section used to calculate/evaluate the rmse

Are you willing to work on this yourself? Sorry, no, I would probably ruin the whole code…

[Optimox]

no it does not have to be an official loss from pytorch, but it does have to be written in pytorch, that’s the only constraint.

[Optimox]

@Carl-McBride-Ellis “I am also sure that the score achieved by TabNet will also improve” : please note that adding rmsle as default evaluation metric will help for ealrly stopping, but won’t change anything in the results of the algorithm itself (will simply change the best epoch). The evaluation metric is not the loss function, but you can pass any loss function to tabnet by providing loss_fn in fit params - this would probably help.