Paths playground

[flying-sheep]

not a PR, just a place to comment and play around before this gets merged into AnnData or scanpy.

Done:

• both obs_df and get_obs_vector support paths as shorthand "X/Actb" or long ("obs", "A/B")
• obs_df returns a Data Frame with unique column names. Smartly resolves collisions: "obsm/X_pca/0" becomes "X_pca1" while "obsm/protein/CD11b_TotalSeqB" becomes "CD11b_TotalSeqB"

TODO:

• allow leaving things out when not ambiguius ("X_pca/0" should work)
• allow sequences: [("obsm", "X_pca", [0, 1])] would expand to [("obsm", "X_pca", 0), ("obsm", "X_pca", 1)]
• obsp support

[review-notebook-app]

Check out this pull request on 

You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.

[ivirshup]

Watch out for this, since column names in dataframes can currently have a "/" in them.

Also, we don't stop dataframe indices from having slashes. We're only expecting two "/" unless the value is in uns, right?

[flying-sheep]

So if we use slashes we need actual parsing then? I mean things are pretty deterministic, so we could set the limit for splitting depending on the attribute. That would be super headachy when combined with fallbacks and shorthand though.

[ivirshup]

If I'm thinking about the same thing, I think we always needed actual parsing. Dataframe indices and columns can be arbitrary strings.

[ivirshup]

I don't think we should allow access strings like: "obsm/X_pca/0", since the integer is ambiguous for dataframes. Is it meant to be the position, or the value?

[flying-sheep]

that’s why I didn’t want data frames in .obsm.

It’s unambiguous when X_pca is an array, and named stuff should go into dedicated anndata modes (which don’t exist yet)

we can fix that in AnnData by not allowing data frames with int colnames or colnames that look like ints…

[ivirshup]

that’s why I didn’t want data frames in .obsm

They're so useful though. I think it'd be fine to just have something like ad.Ref('obsm/X_pca", 0)

[flying-sheep]

That’s just inconsistent and more verbose than a tuple. Why separate the first descent by a slash in a string and the second by a comma?

This seems simpler: ("obsm", "X_pca", 0)

Or this if we have an array: "obsm/X_pca/0"

[ivirshup]

I was thinking more separating the container and the index. This could also allow:

ad.Ref("obsm/X_pca", (0, 1, 2))
ad.Ref("obs", r"leiden*")

Of course, the init could have a signature like ___init__(*args) and pass off to a recursive parsing function.

[flying-sheep]

What’s the difference between a key and an index? is thing in obs[thing] a key or an index? I’d say it’s both.

We used to call that index “component”, but that falls flat when using a data frame and specifying a column name …

[Zethson]

Given that this includes the citeseq tutorials I don't think that this is relevant anymore