Don't call to_output on a cupy array

[Matt711]

• Closes [BUG] AttributeError when combining OneHotEncoder and ColumnTransformer #7039
  AFAIK

  cuml/python/cuml/cuml/_thirdparty/sklearn/preprocessing/_column_transformer.py

  Lines 292 to 297 in 242829b

  	def _transform_one(transformer, X, y, weight, **fit_params):
  	res = transformer.transform(X).to_output('cupy')
  	# if we have a weight for this transformer, multiply output
  	if weight is None:
  	return res
  	return res * weight

  cuml/python/cuml/cuml/preprocessing/encoders.py

  Lines 443 to 447 in 242829b

  	if not self.sparse_output:
  	ohe = ohe.toarray()
  	return ohe

  the call to transformer.transform(X) returns a cupy array, so we shouldn't need to call to_output('cupy') on it.

[Matt711]

If this is the right fix, where should I put the appropriate unit test?

[jcrist]

Tests for ColumnTransformer are in cuml/tests/test_compose.py, that's where I'd add a test for this.

[jcrist]

Thanks for the PR! I think we can do a simpler fix here (as documented below), but we definitely should do some follow-up work to make cuml.preprocessing estimators better follow our type reflecting conventions.

[jcrist]

There are two issues here:

• OneHotEncoder doesn't follow our type reflection conventions (I think much ofcuml.preprocessing also has this issue).
• ColumnTransformer needs to be more robust to consuming disparate output types from transformers (if nothing else, user defined transformers may return a variety of types).

I think the best immediate fix for this issue is something like:

with cuml.using_output_type("cupy"):
    return transformer.transform(X)

This should ensure that any properly implemented cuml transformer returns a cupy array (and not a CumlArray or something else). This is already done for _fit_transform_one below, but not for _transform_one here.

The consumer of this output then handles possibly sparse outputs so we don't want to error here if the return type is somethign else. So the branching structure here in _transform_one isn't needed.

[Matt711]

Thanks! Do you mind filing an issue for the follow-up work?

[jcrist]

Tests for ColumnTransformer are in cuml/tests/test_compose.py, that's where I'd add a test for this.

[yinchi]

Hi, original submitter of #7039 here.

Given that the last two commits were simply merging from branch-25.10 and the failure from the last attempt was a totally unrelated timeout issue, is it possible to try again?

[jcrist]

Apologies for dropping this earlier. I've pushed a small test improvement, this LGTM!

[jcrist]

/merge

[csadorf]

Blocked by #7230 .

[csadorf]

@Matt711 I am very sorry that it took us so long to be able to accept and merge your contribution. Thank you very much for your patience!