TabEBM broken with recent TabPFN + test broken.

[YFrisch]

Describe the bug

from tabpfn_extensions.tabebm.tabebm import TabEBM

throws

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[16], line 1
----> 1 from tabpfn_extensions.tabebm.tabebm import TabEBM
      3 print("\n" + "=" * 50)
      4 print("GENERATING SYNTHETIC DATA WITH TabEBM")

File ~/.../site-packages/tabpfn_extensions/tabebm/tabebm.py:16
     13 from tabpfn_common_utils.telemetry import set_extension
     14 from torch.utils.data import DataLoader
---> 16 from tabpfn.config import ModelInterfaceConfig, PreprocessorConfig
     17 from tabpfn.utils import meta_dataset_collator
     18 from tabpfn_extensions.utils import TabPFNClassifier

ModuleNotFoundError: No module named 'tabpfn.config'

tabpfn==6.0.6
tabpfn-extensions @ git+https://github.com/PriorLabs/tabpfn-extensions.git@7a900321c6aaea8e58ae23117628eaf18cc5a4ab

Steps/Code to Reproduce

No response

Expected Results

No response

Actual Results

No response

Versions

[oscarkey]

hey @YFrisch , I noticed you closed this but as far as I can tell it's still an issue?

It seems like it wasn't picked up because the tests catch import error and use this to decide not to test TabEBM:

tabpfn-extensions/tests/test_tabebm.py

Line 21 in 1960cc6

except ImportError:

so we should fix that too.

[klemens-floege]

@SilenceX12138 would you mind taking a look into this? thanks a lot :)

[SilenceX12138]

Hi @klemens-floege. Thank you for pinning this.

I have carefully checked the latest APIs from TabPFN, and I noticed that there seems to be a major update on the data preprocessing pipeline. As TabEBM relies heavily on the preprocessing pipeline for differentiable inputs, this means the aforementioned TabPFN updates may lead to a drastic break and refactoring of the current TabEBM implementations.

For now, I would highly suggest that users roll back to the compatible TabPFN versions for deploying TabEBM, which is precisely specified in the requirement configs.

[klemens-floege]

@SilenceX12138 thank you for your quick response and answer. There have been some major refactors in the PreProcessing pipeline indeed and our cross package tests did not catch that, sorry about that!; @YFrisch please see the comment above. I think this would resolve the issue for now; we can still leave it open since it would be great to reconcile this in the future.