How to use TabPFN as a module in a larger model

Describe the workflow you want to enable

Hello TabPFN community,

I am trying to apply the pretrained TabPFN as a block inside a larger model, with the following flow structure. The Encoder and Decoder will be learned duruing backpropagation, while the TabPFN is fixed and the gradient information passes through TabPFN.

Input --> Encoder --> TabPFN --> Decoder --> Output

In such case, the TabPFN.get_preprocessed_datasets() gets the tensor matrix from Encoder as the input containing gradient information. However, current function only accepts Numpy as the input for preprocessing. I am wondering whether there is any quick solution for this case, enventhough it was mentioned in the function initialize_dataset_preprocessing that the differentiable input is not supported for regressors yet.

Describe your proposed solution

A new pipeline to allow TabPFN to work as a part in the larger modela and pass the gradient computation graph

Describe alternatives you've considered, if relevant

No response

Additional context

No response

Impact

None