Refinements to the schema for `MultiscaleSpatialImage`

[LucaMarconato]

Here is how a MultiscaleSpatialImage looks like

DataTree('None', parent=None)
├── DataTree('scale0')
│ Dimensions: (c: 3, y: 64, x: 64)
│ Dimensions without coordinates: c, y, x
│ Data variables:
│ image2d_multiscale (c, y, x) float64 dask.array<chunksize=(3, 64, 64), meta=np.ndarray>
└── DataTree('scale1')
 Dimensions: (c: 3, y: 32, x: 32)
 Dimensions without coordinates: c, y, x
 Data variables:
 image2d_multiscale (c, y, x) float64 dask.array<chunksize=(3, 32, 32), meta=np.ndarray>

and this is an example of the second level of the multiscale

data['scale1']['image2d_multiscale']
Out[4]: 
<xarray.DataArray 'image2d_multiscale' (c: 3, y: 32, x: 32)>
dask.array<mean_agg-aggregate, shape=(3, 32, 32), dtype=float64, chunksize=(3, 32, 32), chunktype=numpy.ndarray>
Dimensions without coordinates: c, y, x
Attributes:
 transform: Sequence \n Scale (y, x)\n [2. 2.]\n Identity 

We need to add the following constraints to the validator:

1. the nodes of the tree are only 'scale0', 'scale1', etc.
2. each node has only one data variable (image2d_multiscale in the example above). The name is irrelevant
3. each data variable has the transformation in .attrs['transform']. That is, the transformation is not here: multiscale_data.attrs['transform']; it's not here: multiscale_data.attrs['scale0'].attrs['transform'], but it's here: multiscale_data.attrs['scale0']['image2d_multiscale'].attrs['transform'].

Right now this type of validation is scattered around the repo. In particular this is done in write.py, in the function _iter_multiscale(), here reported

def _iter_multiscale(
 data: MultiscaleSpatialImage,
 attr: str,
 key: Optional[str] = None,
) -> list[Any]:
 # TODO: put this check also in the validator for raster multiscales
 name = None
 for i in data.keys():
 variables = list(data[i].variables)
 if len(variables) != 1:
 raise ValueError("MultiscaleSpatialImage must have exactly one variable (the variable name is arbitrary)")
 if name is not None:
 if name != variables[0]:
 raise ValueError("MultiscaleSpatialImage must have the same variable name across all levels")
 name = variables[0]
 if key is None:
 return [getattr(data[i][name], attr) for i in data.keys()]
 else:
 return [getattr(data[i][name], attr).get(key) for i in data.keys()]

and it is done in get_transform() from core_utils.py:

 d = dict(e['scale0'])
 assert len(d) == 1
 xdata = d.values().__iter__().__next__()
 t = get_transform(xdata)

[LucaMarconato]

For the moment I remove completely the attrs constraint in the validator (RasterSchema). In fact this would be valid only for SpatialImage, not for MultiscaleSpatialImage.

[giovp]

agree on everything. I will wait for transform to get merged and update it.

[giovp]

1. the nodes of the tree are only 'scale0', 'scale1', etc.

done

2. each node has only one data variable (image2d_multiscale in the example above). The name is irrelevant

done

3. each data variable has the transformation in .attrs['transform']. That is, the transformation is not here: multiscale_data.attrs['transform']; it's not here: multiscale_data.attrs['scale0'].attrs['transform'], but it's here: multiscale_data.attrs['scale0']['image2d_multiscale'].attrs['transform'].

I think this can still be validated with xarray-schema, will check and remove the comment if so.