Adding data variables without specifying nan values explicitly? Only via a list of dimensions?

[doronbehar]

Is your feature request related to a problem?

Very often, I find it much more comfortable when creating complex xarray.Datasets, to create data variables in the following way (also implemented as a workaround just below):

1. Declare the dimensions, each of which is defined by a 1D array (of varying sizes of course)
2. Add the data variables based on the sizes of the dimensions the xarray.Dataset is already aware of.

This way, I don't have to manually construct np.full arrays that would fit my data - I only have to specify the dimensions as a list of strings, and the shape is already constructed.

Describe the solution you'd like

I use the following:

def set_dvar_from_dims(self, name, specific_dims, astype=np.float64, **kwargs):
    for kw in ['dims', 'coords']:
        assert kw not in kwargs, (
            f"Don't specify {kw} when using set_data_variable_from_dims, this "
            "doesn't make sense"
        )
    self[name] = xr.DataArray(
        dims=specific_dims,
        coords={dim: self[dim] for dim in specific_dims},
        **kwargs,
    ).astype(astype)
xr.Dataset.set_dvar_from_dims = set_dvar_from_dims

And example usage:

ds = xr.Dataset(
    coords=TEMPERATURES_SOURCES | {
        'mass': self.masses,
        'regime': TEMPERATURES_REGIMES,
        'break_idx': range(TEMPERATURES_REGIMES_AMOUNT + 1), 
        'ns': list("ns"),
        'time|cooling': times,
    },
)
common_dims = ['mass'] + list(TEMPERATURES_SOURCES.keys())
ds.set_dvar_from_dims('coolFreqBreak', common_dims)
ds.set_dvar_from_dims('beta', common_dims + ['break_idx'])
ds.set_dvar_from_dims('coolFreq', common_dims + ['regime', 'ns'])
ds.set_dvar_from_dims("fit_prediction", common_dims + ["time|cooling"])

Describe alternatives you've considered

No response

Additional context

No response

[TomNicholas]

This way, I don't have to manually construct np.full arrays that would fit my data

Would the existing xr.full_like followed by expand_dims not work for you?

[doronbehar]

This way, I don't have to manually construct np.full arrays that would fit my data

Would the existing xr.full_like followed by expand_dims not work for you?

Well indeed using xr.full_like might help specifically for not using np.full, but I don't see how I can avoid specifying the dimensions and their values for every data variable I want to add.

[TomNicholas]

Using full_like or empty_like seems to allow you to avoid ever specifying values, correct?

And you can add multiple variables then expand all their dims at once with ds.expand_dims. And you can pass any common sizes you want on new dims in at that point. So doesn't that allow you to avoid specifying dimensions for each new variable separately?

So I guess I don't quite yet understand what this new API would allow you to do that cannot be achieved relatively easily with existing API.

[doronbehar]

Using full_like or empty_like seems to allow you to avoid ever specifying values, correct?

It helps me avoid specify the values of the data variables, but not the values of the dimensions, or their names.

In fact, if I understand you correctly, what you are suggesting doesn't even require xr.full_like, if we are interested in np.nan defaults, as xarray.DataArray defaults to np.nan anyway, and that's what I'm looking for as default values.

More over, I think I managed to translate your suggestion to my example:

# Create dummy DataArrays and use zeros_like
ds['coolFreqBreak'] = xr.DataArray(
    dims=common_dims,
    coords={dim: ds.coords[dim] for dim in common_dims}
)
ds['beta'] = xr.DataArray(
    dims=common_dims + ['break_idx'],
    coords={dim: ds.coords[dim] for dim in common_dims + ['break_idx']}
)
ds['coolFreq'] = xr.DataArray(
    dims=common_dims + ['regime', 'ns'],
    coords={dim: ds.coords[dim] for dim in common_dims + ['regime', 'ns']}
)
ds["fit_prediction"] = xr.DataArray(
    dims=common_dims + ["time|cooling"],
    coords={dim: ds.coords[dim] for dim in common_dims + ["time|cooling"]}
)

The tedious part is to specify the dimensions in the dims and coords arguments, not to mention the need to create a dictionary for the coords argument, when all that information is already available in the xarray.Dataset.