search

automl / PFNs

Our maintained PFN repository. Come here to train SOTA PFNs.
Apache License 2.0
34 stars 6 forks source link

Understanding the loss function for FullSupportBarDistribution #3

Closed pg2455 closed 5 months ago

pg2455 commented 5 months ago

Hi,

I have been trying to understand the loss function for the regression case. The code for FullBarDistriution.forward Line 251 has mean_prediction_logits as its parameter.

I couldn't find any reference to this in the main paper. Apologies if I missed it! Could you please explain what is it doing and whether it is necessary to do so? My concern is also regarding the mismatch of the dimensions, i.e., Line 257 require nll_loss and smooth_loss to be of the same dimension but it might not be possible after Line 251.

SamuelGabriel commented 5 months ago

it might just be a naming issue: in the paper we call it "riemann distribution", that is the same as the full bar distribution.

pg2455 commented 5 months ago

This didn't really answer my question, but I am able to make it work without needing any mean_prediction_logits.