De-hack uns for single gene

[csweaver]

data.uns = cell_data.uns fails if there is a single gene. Need to fix.

[csweaver]

This may be a scanpy error, get Alex minimal test case.

[csweaver]

Submitted issue to scanpy: scverse/scanpy#323

[falexwolf]

Fixed in scverse/anndata@ae8a1d0. Thank you!

[bkmartinjr]

FYI I'm also seeing a non-trivial performance cost to the .uns deepcopy() on each slice of an anndata object. Still digging into why, but given that the cellxgene back-end has no dependency on this, we may just want to clobber .uns upon load

The code in question: anndata/base.py:_init_as_view(), lines 689::

        uns_new = deepcopy(self._adata_ref._uns)

[bkmartinjr]

@falexwolf - thanks for the fix! Do you expect it to be released soon? It would allow us to remove a bunch of special-case code which handles the forced dimension drop. Thanks!

[bkmartinjr]

Very related issue, just filed on scanpy -- you can't slice both dimensions if one of the results is single-valued. There appears to be an internal inconsistency with how dimension flattening is done in AnnData.

scverse/scanpy#332

[bkmartinjr]

@csweaver - I will clean up the uns work-around as part of the work I'm doing on slicing performance. It is closely related to the issue I just filed against scanpy.

[falexwolf]

FYI I'm also seeing a non-trivial performance cost to the .uns deepcopy() on each slice of an anndata object. Still digging into why, but given that the cellxgene back-end has no dependency on this, we may just want to clobber .uns upon load

Sorry about the late response: .uns also stores the graph of cells that is used in the embeddings. This graph is sliced when slicing along the observations dimension. It is just copied when slicing along the variables dimension.

I see that this is an efficient solution if you want to extract single genes, etc.

Maybe one should find a more light-weight solution for that? But I'd need to know a bit more...

[bkmartinjr]

wrt .uns copy time -- our current use case boils down to slicing, on both axis using boolean array selectors, to access either annotations or X. We do not (yet) expose/use the info stored in .uns to the visualization.

So the "need" is to slice and access X or annotations, without the overhead of copying .uns.

I'm not sure this is a high priority item, especially as it may require API changes, and isn't our largest time sink. Hence the air quotes.

[falexwolf]

We should have a better way for serving this case in the future, agreed!