Thomas J. Fan

> Whether that's the case in scikit-learns transformers, I don't know... By default, scikit-learn estimators do not write to the underlying buffer. There are common tests to make sure that...

> (but just out of interest, why would it require a hard dependency?) You are right, polars can be a soft dependency. If we get a DataFrame, then the behavior...

> As for my perspective with writing tools on top of sklearn, it's pandas all the way right now. If there's any nuances with respect to user input into sklearn...

For record keeping, this looks like a random seed that failed: ```python global_random_seed = 34 def test_minibatch_sensible_reassign(global_random_seed): # check that identical initial clusters are reassigned # also a regression test...

@adrinjalali @lorentzenchr This PR is ready to review. With the recent updates, I simplified the PR a bit and reduced the diff.

`sklearn.utils._array_api.isdtype` was not designed to be used with NumPy directly. It was designed to be only used through `_NumPyAPIWrapper` or `_ArrayAPIWrapper`. When using the `_NumPyAPIWrapper`, the types are converted to...

In any case, I am okay with vendoring the implementation from `array_api_compat.xp.isdtype`.

Implementation wise, I think we follow the method described in 3.3, and **do not** drop any of the classes. Dropping any of the classes will introduce a bias for downstream...

> Why do we need to vendor lazy_loader ? can't we make it a dependency ? Given how conservative we are with adding new dependencies, I decided to vendor the...

@rth I just saw your EuroPython slides, do you think using this PR's lazy import mechanism will help with load times on WASM?