predict only with tabPFN

[viotemp1]

Hello,
I want to decouple model (classifier) creation and fit from model predict in two different scripts.
Is it possible to have?

1. model creation, fit and save model
2. load trained model in step 1 and predict
   Thanks

[iivalchev]

Hi there, see #292, might help.

[viotemp1]

thanks!

[carbocation]

Hi there, see #292, might help.

This seems prohibitive for medical uses (because you'd need to bundle your training data into the package). Presumably fine tuning will ultimately be the only acceptable approach. Although, contrary to this, it looks like save_fitted_tabpfn_model may accomplish the goal?

[priorphil]

Even with fine-tuning you'd likely be able to recover the training data with some work (the same holds for most other learning algorithms, see e.g. for a random forest reconstruction https://arxiv.org/html/2402.19232v1) :)