search

jku-vds-lab / paradime

A framework for parametric dimensionality reduction
MIT License
9 stars 3 forks source link

Implement exaggeration #14

Closed einbandi closed 2 years ago

einbandi commented 2 years ago

Implement (early) exaggeration, which means that relation values are rescaled by a given factor for the first few epochs. The relation loss has to be rescaled accordingly. This could be specified in the relation loss.

einbandi commented 2 years ago

Much easier, and a nice proof of the generalization of training phases: implement Exaggerate transform that simply multiplies the relations by a given factor. Still, to get correct scaling of the relation loss it will be necessary to detect the existence of an exaggeration transform inside ParametricDR and pass the info on to the loss.

einbandi commented 2 years ago

I decided to not implement this at all for now. Users can specify exaggeration easily via a Functional transform. The just have to rescale the loss themselves, if at all necessary.