Method to retrieve weights as a DataArray

[aulemahal]

While coding the spatial averaging in #24, I realized that the Regridder object was missing an option to retrieve a usable array of the regridding weights. While this is not common when performing conventional regridding, I think that most users doing spatial averaging would like to see or save the weights masks.

The Regridder object could have a simple get_weights method that returns a DataArray of the expanded sparse matrix with 4 or 3 dimensions. Dimensions would be renamed with "_in" and "_out" suffixes.

[raphaeldussin]

my understanding is that ESMF only produces 2d regridding matrices (I could be wrong).
The size of the sparse matrix being (nx_src * ny_src) x (nx_dst * ny_dst), are you suggesting something like:

def get_weights(regridder):
    return xr.DataArray(regridder.weights.toarray(), dims=('nxy_src', 'nxy_src'))

[aulemahal]

I would suggest something a bit more complex that outputs an array of shape (lon_in, lat_in, lon_out, lat_out) or (lon_in, lat_in, locations) in a locstream/polylist_out case. This adds a reshaping operation, but shouldn't be difficult.

[huard]

@aulemahal Could you repurpose or generalize smm.read_weights ?

[aulemahal]

I'm not sure how that would serve the purpose. May be the issue is not clear, I am suggesting pretty much this:

def get_weights(self):
    if self.locstream_in:
        dims = ['locations_in']
        shape = [self.n_in]
    else:
        dims = [dim + '_in' for dim in self.input_horiz_dims]
         shape = list(self.shape_in)

    if self.locstream_out:
        dims.append('locations_out)]
        shape.append(self.n_out)
    else:
        dims.extend([dim  + '_out' for dim in self.out_horiz_dims])]
         shape.extend(self.shape_out)
    return xr.DataArray(self.weights.toarray().reshape(shape), dims=dims)

This would also need the new attribute input_horiz_dims to be set in the __init__, it could still be overridden in the regrid methods anyway.

[raphaeldussin]

@aulemahal I'm not sure that would work for large arrays. Say if I regrid a 1/4 degree regular grid onto a 1 degree (pretty standard stuff), a 4d matrix would be of size 1440x720x360x180 = 67e9 elements = 269 GB for single precision

[aulemahal]

You're right... I was thinking of the cases using polygons, where the weights could actually be useful themselves. This could be a small example in the doc instead. So, if the user explodes their ram, it's their own fault.

[aulemahal]

Just remembered this issue. It was "solved" by an example in the doc of #24.