Hi @adam2392,

I am following up with this PR which I forgot.

I do not have commit rights on @neurodata's fork of scikit-learn.
Can you merge main in the branch of this PR?

I do not have commit rights on @neurodata's fork of scikit-learn. Can you merge main in the branch of this PR?

Done!

I think it is worth adding float32 support in tree based models before working further on extending tests as data is cast to float64 at the moment:

https://github.com/scikit-learn/scikit-learn/blob/ba5f9e0cd10486f954299a7796c862dd0ddc9470/sklearn/tree/_tree.pyx##L101-L102

I think it is worth adding float32 support in tree based models before working further on extending tests as data is cast to float64 at the moment:
https://github.com/scikit-learn/scikit-learn/blob/ba5f9e0cd10486f954299a7796c862dd0ddc9470/sklearn/tree/_tree.pyx##L101-L102

What's the difficulty here? Could we just re-define the trees and sub the data types for a fused datatype? If so, I think I can submit a straightforward PR. Lmk if you think this is insufficient tho.

E.g.

ctypedef cnp.npy_float32 DTYPE_t          # Type of X
ctypedef cnp.npy_float64 DOUBLE_t       # Type of y, sample_weight

ctypedef fused NEW_DTYPE:
   DTYPE_t
   DOUBLE_t

Unfortunately fused-types and Cython extension types can't be used simultaneously.

More information and resolution pathways are given here: #7059 (comment)

Oh yeah forgot bout that unfortunate limitation :p

Things would significantly be simpler for scikit-learn if Cython were to support it.

IIRC, supporting this really is complicated because one has to deal with many historical aspects and choices of Pyrex which Cython supersedes.

Edit: cython/cython#3283 tracks discussions on this topic.