ENH : add_parameter to keep dim in TransformedTargetRegressor #30363
Conversation
|
@glemaitre What do you think about this? |
|
From one of the comments #26530 (comment) it sounds like we change the number of dimensions. That feels like a bug to me -> can't we fix this without having to introduce a new parameter? I'm not sure about the answer to that question, just started looking into this. But maybe someone already knows? |
There was a problem hiding this comment.
Thanks for the PR @Irdasset
Instead of adding a new parameter could you handle this by checking the original dimension of y as discussed in the comments on the issue?
|
Hi I had a question regarding the current behavior of TransformedTargetRegressor. From what I understand, even when a 2D target y (e.g., shape (n_samples, 1)) is passed, the inverse transform ends up returning a squeezed 1D array. This breaks compatibility with estimators that strictly expect 2D output (e.g., skorch, custom wrappers, etc.). I saw in #26530 and #30363 that you're discussing ways to preserve the shape or add a parameter like preserve_y_shape. 👉 Is this still open for contributions or test case additions? I’d like to help or maybe try adding a patch if needed. Thanks again! |
|
The solution that was adopted was to not introduce a parameter and directly fix the behavior because it was a bug. It has been fixed by #31563, so we can close this PR. Thanks for the contributions. |
Reference Issues/PRs
Fixes #26530
What does this implement/fix? Explain your changes.
This implement allow the user to preserve the output dim by adding a parameter
preserve_y_dim