feature request: @datasetclass

[shaunc]

xarray-custom is a great idea -- thanks for the project!

What I was actually looking for was a way to provide this sort of functionality for a dataset, rather than a dataarray. Is this on the roadmap?

[astropenguin]

@shaunc I am sorry for the late reply. Unfortunately, I do not plan to implement @datasetclass to xarray-custom because it is only optimized for DataArray creation. But, because I also think that it is useful, I made another package that provides both @dataarrayclass and @datasetclass.

astropenguin/xarray-dataclasses: xarray extension for DataArray and Dataset classes

Please note that creating custom method is out of scope in this package. I hope this would be something useful for you...

[shaunc]

Great -- thank you! I'll check it out!

Currently I'm using a hand-rolled schema checker in test code. A few questions:

1. I don't need quite "production ready" ... but nice to have "beta" quality ("feature complete wrt to a stable version, test coverage good; may have some bugs, but should be relatively straightforward to fix"). How near is array-dataclasses wrt to being "beta quality"?
2. Currently I often use xr.open_dataset(path, engine="netcdf4") to create a dataset. Is there a recipe for creating a datasetclass from an external source -- or would it be easy to add? (E.g. can I add a classmethod ... hmm... does new work with a passed in dataset?)

[astropenguin]

How near is array-dataclasses wrt to being "beta quality"?

I think it is close to beta: I do not plan to change the coding style of user side.
But, as you already know, we may add new features with keeping backward compatibility.

Is there a recipe for creating a datasetclass from an external source -- or would it be easy to add?

For the purpose of schema check, adding a custom class method would be useful:

import numpy as np
import xarray as xr
from xarray_dataclasses import Data, datasetclass


@datasetclass
class RGBImage:
    """DataArray that represents RGB images."""

    red: Data[tuple["x", "y"], float]
    green: Data[tuple["x", "y"], float]
    blue: Data[tuple["x", "y"], float]

    @classmethod
    def from_dataset(cls, dataset: xr.Dataset) -> xr.Dataset:
        """Try to create a Dataset instance from another one."""
        return cls.new(
            **dataset.data_vars,
            **dataset.coords,
            **dataset.attrs,
        )


@datasetclass
class CMYKImage:
    """DataArray that represents CMYK images."""

    cyan: Data[tuple["x", "y"], float]
    magenta: Data[tuple["x", "y"], float]
    yellow: Data[tuple["x", "y"], float]
    black: Data[tuple["x", "y"], float]

Then you can check if a dataset is compatible with a definition:

shape = 5, 5


rgb_image = RGBImage.new(
    np.ones(shape),  # red
    np.ones(shape),  # green
    np.ones(shape),  # blue
)


cmyk_image = CMYKImage.new(
    np.ones(shape),  # cyan
    np.ones(shape),  # magenta
    np.ones(shape),  # yellow
    np.ones(shape),  # black
)


RGBImage.from_dataset(rgb_image)  # ok
RGBImage.from_dataset(cmyk_image)  # TypeError