Can not write array and dgeMatrix objects

[ToryDeng]

I am unable to write objects of type array and dgeMatrix to .h5ad files. Here is an example R script demonstrating this issue:

adata <- generate_dataset(format = "AnnData")
adata$uns <- list(arr=array(c(1:30),dim = c(3,2,5)))
write_h5ad(adata, "test_adata.h5ad")
warnings()

Warning messages:
1: In value[[3L]](cond) :
  Could not write element 'layers/numeric_dense' of type 'dgeMatrix':
unknown type
2: In value[[3L]](cond) :
  Could not write element 'layers/numeric_dense_with_nas' of type 'dgeMatrix':
unknown type
3: In value[[3L]](cond) :
  Could not write element 'layers/integer_dense' of type 'dgeMatrix':
unknown type
4: In value[[3L]](cond) :
  Could not write element 'layers/integer_dense_with_nas' of type 'dgeMatrix':
unknown type
5: In value[[3L]](cond) :
  Could not write element 'obsm/numeric_dense' of type 'dgeMatrix':
unknown type
6: In value[[3L]](cond) :
  Could not write element 'obsm/numeric_dense_with_nas' of type 'dgeMatrix':
unknown type
7: In value[[3L]](cond) :
  Could not write element 'obsm/integer_dense' of type 'dgeMatrix':
unknown type
8: In value[[3L]](cond) :
  Could not write element 'obsm/integer_dense_with_nas' of type 'dgeMatrix':
unknown type
9: In value[[3L]](cond) :
  Could not write element 'varm/numeric_dense' of type 'dgeMatrix':
unknown type
10: In value[[3L]](cond) :
  Could not write element 'varm/numeric_dense_with_nas' of type 'dgeMatrix':
unknown type
11: In value[[3L]](cond) :
  Could not write element 'varm/integer_dense' of type 'dgeMatrix':
unknown type
12: In value[[3L]](cond) :
  Could not write element 'varm/integer_dense_with_nas' of type 'dgeMatrix':
unknown type
13: In value[[3L]](cond) :
  Could not write element 'obsp/numeric_dense' of type 'dgeMatrix':
unknown type
14: In value[[3L]](cond) :
  Could not write element 'obsp/numeric_dense_with_nas' of type 'dgeMatrix':
unknown type
15: In value[[3L]](cond) :
  Could not write element 'obsp/integer_dense' of type 'dgeMatrix':
unknown type
16: In value[[3L]](cond) :
  Could not write element 'obsp/integer_dense_with_nas' of type 'dgeMatrix':
unknown type
17: In value[[3L]](cond) :
  Could not write element 'varp/numeric_dense' of type 'dgeMatrix':
unknown type
18: In value[[3L]](cond) :
  Could not write element 'varp/numeric_dense_with_nas' of type 'dgeMatrix':
unknown type
19: In value[[3L]](cond) :
  Could not write element 'varp/integer_dense' of type 'dgeMatrix':
unknown type
20: In value[[3L]](cond) :
  Could not write element 'varp/integer_dense_with_nas' of type 'dgeMatrix':
unknown type
21: In value[[3L]](cond) : Could not write element 'uns/arr' of type 'array':
argument is not a matrix

Is there a way to resolve this issue, or is this functionality not yet implemented in anndataR?

[lazappi]

We don't currently have support for either of these formats. Maybe we should consider it but neither of them is currently used. I'm not sure that generate_dataset() should output dgeMatrix objects rather than normal dense matrices and I'm a bit surprised it hasn't caused issues before now. Possibly something changed in {Matrix} or another package?

I'm less sure about supporting array as I'm not sure the H5AD specification has a definition for how to store n-dimensional arrays.

@LouiseDck @rcannood Do you have any other thoughts?

[LouiseDck]

I was running into these dgeMatrix issues as well and was trying to figure out when you would want to use them instead of normal dense matrices? If I understand, you just wouldn't?

As far as I could figure out, there's no real difference between a matrix and an array object?

The generate_dataset() functionality is also there to systematically evaluate (in the tests I'm writing atm) which R datatypes we should write to h5ad and how they compare to the Python datatypes. One of the goals of this is to have a list of supported an unsupported datatypes.

[lazappi]

I was running into these dgeMatrix issues as well and was trying to figure out when you would want to use them instead of normal dense matrices? If I understand, you just wouldn't?

Yeah, I don't think there is any reason to use a dgeMatrix, I wasn't even aware they existed until now. If the default arguments to the dummy function create them then I think they should be handled in a better way, otherwise users will get confused why things are failing.

As far as I could figure out, there's no real difference between a matrix and an array object?

A matrix is always 2D but an array is n-dimensional. In most cases a matrix is what you want but #204 was just opened so array probably needs to be supported as well (for slots where dimensions != 2 are allowed).

[rcannood]

Let's fix this. It's true that a regular matrix could be used instead of a dgeMatrix, but a dgeMatrix is still a thing so HDF5AnnData should not flip out when we pass a dgeMatrix to it ^^