TabPFN for Federated Learning

[SlamanigG]

Hello! 👋

I am currently trying to use TabPFN in a FL system using Flower. For my use case I need to be able to specify the number of local rounds, i.e. train the TabPFN like a regular PyTorch model for a set number of epochs.

It seems like this is possible using the fine-tuning example, which calls model.fit_from_preprocessed() instead of model.fit(). However, I somehow could not find a way to evaluate the trained model on a test or validation set, without calling model.fit() first, because of this error: AssertionError: Invalid forward pass: Bad combination of inference mode, input X, or executor type. Ensure call is from standard predict or a batched fine-tuning context.

This causes issues when evaluating an aggregated model from the server on the individual clients, as i do not want to re-fit the model here but rather just receive it from the server and evaluate it on my test data.

Is there any way to train the TabPFN for a set amount of epochs and then evaluate it on some test or validation set?