Fix: Long vs Float dtype mismatch in TabrModel

[miriambt]

Fixed an error in TabrModel when LongTensor is passed into a Linear layer expecting FloatTensor:

RuntimeError: mat1 and mat2 must have the same dtype, but got Long and Float.

This occurs in the forward() method of TabrModel at the line:

context_y_emb = self.label_encoder(candidate_y[context_idx][..., None])

Here, candidate_y[...] is a LongTensor, which causes a type mismatch when passed into nn.Linear, which expects a FloatTensor.

[dholzmueller]

Thanks for the PR! How did you manage to run into this error? For me the old code works just fine, but I don't know why.

[miriambt]

I was running TabM_D_Classifier when the error occurred, and changing the input to .float() resolved it. However, while using TabR_S_D_Classifier today, I ran into the following error:

RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.FloatTensor instead (while checking arguments for embedding)

I believe I have now implemented a fix that handles both TabM_D_Classifier and TabR_S_D_Classifier correctly (see last commit). Before merging the changes, though, it would be great if you could test it with your data as well — just to make sure the fix does not resolve one issue while introducing another. Thanks!

[dholzmueller]

TabM should not be using this code at all. Can you provide a minimal reproducible example that triggers an error with the old code?