Viewing array with a different axes order

[John-P]

I am frequently coming across situations where I would like to view zarr arrays with the axes in a new order. I can do this with numpy using np.moveaxis to get a view. Is this possible with zarr? It looks like I could do this with dask but it seem silly to add a bunch of extra dependencies for a simple alternative view into the data like this. Any advice? Is this already possible or something that could be added if not?

[rabernat]

This is pretty straightforward to do if you use Zarr together with Xarray. See https://docs.xarray.dev/en/stable/user-guide/io.html#zarr for details about how to load Zarr data from Xarray.

[John-P]

Thank you for the suggestion. I have looked into xarray but found that it is very finicky to use with a Zarr array. It can only load a group and not simply an array. Perhaps I can just wrap an array in a group before passing to xarray? Maybe I just don't understand the xarray semantics but I found it very difficult to use with a zarr array. It can also be quite noisy with warnings if the Zarr does not have the non-standard Zarr metadata required (_ARRAY_DIMENSIONS).

It may be ok but just seems like an inelegant solution and adds another dependency to projects where I would like to reduce the number of unnecessary dependencies. Particularly with the recent pypi supply chain attacks.

[John-P]

I realise it might be more helpful to explain why I am not finding xarray as useful as I had hoped. For example, if I have a Zarr array and want to turn it into an xarray dataset I can do this as follows:

import zarr
import xarray as xr

array = zarr.open("large-array.zarr")
dataset = xr.DataArray(array, dims=("t", "y", "x", "c"))

However, this loads the whole zarr array in to memory. I see that xarray has a way to open a zarr group as a dataset but not a zarr array without loading it all into memory. Additionally, the zarr must have special additional metadata for the dimensions. There does not appear to be a way to specify the dimensions when creating a Dataset if this is missing. The datasets which I have do not have this special metadata and so this is a big pain.

[John-P]

I would also like to be able to use zarr stores which are not on disk and where I cannot always point array to a directory (e.g. a tifffile zarr store) but would like to be able to use a zarr store object instead.

[John-P]

The best solution I have found so far is to use dask e.g.:

import tifffile
import dask.array

with tifffile.imread('temp.ome.tif', aszarr=True) as store:
    array = dask.array.from_zarr(store, 0)
array = dask.array.moveaxis(array, (0, 1, 2), (2, 0, 1))

It just seems like something that should be doable with zarr as kind of view. I may go with this for now but it would be nice to not add dask and a lot of other child dependencies for this.