Allowing `.X` to be None

[ivirshup]

Tracking issue for allowing X to be None.

When parts get ticked off they should be marked with the PR that fixed them.

TODO:

All of these need tests, some of them may already work.

• Getter, setter, deleter
• Initializer (seems to work, needs testing)
• Views (basically just check that .X = None doesn't change other view behavior, a[idx].X doesn't do anything weird, and copying views still works)
• _subset_inplace (25bbf8a)
• Copying Improve support for X=none #677
• Figure out inference of AnnData shape when X is None
• Backed mode
• Concatenation
  • When all values of X are None Improve support for X=none #677
  • When some values of X are None concat when some values of X are None #678
• Transpose
• Methods which default to values from X, e.g. obs_vector, to_df. (Improve support for X=none #677)
• IO (seems to work, needs testing)
  • Main writers (zarr, hdf5) Improve support for X=none #677
  • Main readers Improve support for X=none #677
  • Specialized writers like to_csvs
• Raw (pretty similar set of things to be checked)
  • Initializer
  • Backed
  • Copies
  • Concatenation
  • IO

---

There is a bit of ordering to do here.

• Getting the property and initializer working will be required for most everything else.
• IO should happen before backed mode.
• Backed mode and concatenation are pretty independent.

[flying-sheep]

@rcannood which of the above points have been adressed by your PR?

[rcannood]

To begin:

• Getter, setter, deleter
• Initializer
• Transpose

There are also some basic checks for 'Views', but not everything mentioned above has been taken care of so it's still WIP.

[ivirshup]

I've fixed up many of these cases in #677.

I think .X is None is now in a fairly useable state, and think we're in an alright place to push the rest to the next release.

[ivirshup]

Adding a point to the checklist:

Gotta figure out what the shape of the object is when X is None. Currently we figure out the shape of an AnnData by checking the shapes of X, obs, and var and asserting everything else fits that. What we should probably start doing is resolving the shape from all elements, so an AnnData with just obsm/ varm can be created.

---

An interesting question is what happens if you only provide elements aligned to one axis. E.g. just obsm or obs. Is n_vars unknown or zero?

Is this allowed?

adata = AnnData(obs=obs)

Then what is n_vars. Would this be allowed after?

adata.varm["x"] = np.ones((20, 10))