ENH Add dtype preservation to FeatureAgglomeration

[IvanLauLinTiong]

Reference Issues/PRs

In scope of #11000

What does this implement/fix? Explain your changes.

Implement dtype preservation to FeatureAgglomeration and its relevant tests

Any other comments?

[jeremiedbb]

Thanks for the PR @IvanLauLinTiong ! looks good. I just removed one of the tests because it will be covered by a common test.

[glemaitre]

@jeremiedbb In this case, we don't modify the tree builder algorithm to perform the 32 bits operations?

[IvanLauLinTiong]

Thanks for the PR @IvanLauLinTiong ! looks good. I just removed one of the tests because it will be covered by a common test.

Okay thanks

[IvanLauLinTiong]

@jeremiedbb In this case, we don't modify the tree builder algorithm to perform the 32 bits operations?

@glemaitre @jeremiedbb
Hi, do I need to modify the tree builder algorithm to perform 32 bits operations ?

[IvanLauLinTiong]

@glemaitre @jeremiedbb
Hello, any chances to help me review this? Thanks

[IvanLauLinTiong]

Hi @jeremiedbb, could you advise me if there's any mistakes/or improvement that I still need to do for this PR? I'm happy to learn from mistakes and work more on this if there's a room for improvement.
Thanks
Ivan

[glemaitre]

I discussed the problem with @ogrisel IRL. We agreed that we should also make sure that fit preserve dtype and that every computation should be done in float32.

It means that we need to check that every algorithms in TREE_BUILDERS needs to be modified.
It potentially leads to a large diff. Therefore, it could be easier to open a PR by algorithm and do it step by step. It will also be either to review.

[glemaitre]

Converting into draft until it is ready to be reviewed.

[IvanLauLinTiong]

I discussed the problem with @ogrisel IRL. We agreed that we should also make sure that fit preserve dtype and that every computation should be done in float32.

It means that we need to check that every algorithms in TREE_BUILDERS needs to be modified. It potentially leads to a large diff. Therefore, it could be easier to open a PR by algorithm and do it step by step. It will also be either to review.

Okay noted.