[MRG+1] Allow sparse input to incremental PCA

[scottgigante]

Reference Issues/PRs

Fixes #13957.

What does this implement/fix? Explain your changes.

Implements sparse input for IncrementalPCA. IncrementalPCA is by design suited to accepting sparse input; this allows the input to be sparse, and if it is so, converts the data to dense on a batch-wise basis.

[abhinavsagar]

Looks good to me.

[NicolasHug]

Please add an entry to the change log at doc/whats_new/v*.rst. Like the other entries there, please reference this pull request with :issue: and credit yourself (and other contributors if applicable) with :user:.

[jnothman]

If I want to apply partial_fit batchwise to multiple input matrices, I would not be able to run step 2 without step 1.

yes, but if the way you treat sparse matrices is to densify them, you might as well get the user to do that...

[scottgigante]

If I want to apply partial_fit batchwise to multiple input matrices, I would not be able to run step 2 without step 1.

yes, but if the way you treat sparse matrices is to densify them, you might as well get the user to do that...

A clarification: if the user wants to batchwise densify and fit multiple sparse matrices, then this is not currently possible. Example scenario: you want to fit a single estimator to multiple large mtx files. Alternative solution would be to load all of the files, concatenate the matrices with scipy.sparse.vstack and then apply fit.

[jnothman]

Why can't the user currently pass them, dense, one by one to partial_fit?

[scottgigante]

I would presume that turning the entire sparse matrix to dense at once would be undesirable for memory reasons.

[jnothman]

That's why we require the user to make the data dense in situations where doing it automatically may be deleterious.

[scottgigante]

Note: it's failing tests due to some other unrelated change, I think - failing tests in grid search CV.

[scottgigante]

@NicolasHug @jnothman any further comments?

[jnothman]

Sorry my review time has been limited. Please add a check that partial_fit still raises an appropriate error when passed sparse X. Otherwise lgtm, thanks!

[NicolasHug]

As Joel mentioned please test the error in partial_fit for dense input.

Also all the methods that accept/return a sparse X should be changed to array-like or sparse matrix

Otherwise LGTM too!

[scottgigante]

@jnothman @NicolasHug should be good to go now.

[NicolasHug]

Thanks @scottgigante !

[NicolasHug]

Thanks @scottgigante !

[jnothman]

Thanks Scott!

[scottgigante]

Thanks @jnothman and @NicolasHug for the detailed reviews!