Is there a version accepts tensor vectors directly?

[Rockou]

Hi, great work! I would like to know if there is a version of TabPFN v2 that accepts tensor vectors directly for both input and output (the current version uses NumPy for this). The goal is to enable backward propagation during parameter optimization in TabPFN v2.

Thanks a lot!

[iivalchev]

You can find how to do that here: https://github.com/LennartPurucker/finetune_tabpfn_v2

[Rockou]

You can find how to do that here: https://github.com/LennartPurucker/finetune_tabpfn_v2

Hi, thank you very much! I will try to fix this issue with the model you provided.

[Rockou]

Hi, I have reviewed the model you provided and noticed that it employs the PerFeatureTransformer module as the output of TabPFN v2. To verify its behavior, I compared the responses under two scenarios: one directly using the TabPFN v2 code from the official website, and the other incorporating the PerFeatureTransformer, both with the same input vectors and random seed. Theoretically, if the PerFeatureTransformer module represents the output of TabPFN v2, both cases should yield identical results. However, the responses differ between the two scenarios. This suggests that the PerFeatureTransformer module may not fully represent the output of TabPFN v2. In other words, it seems insufficient to solely use the PerFeatureTransformer module as a replacement for the output of TabPFN v2. I would greatly appreciate it if you could provide some insights or clarification on this issue. Thank you very much.