error with Canvas().raster on xarray 0.17

[darribas]

ALL software version info

The full list is available here, the gist is:

• Python 3.8
• xarray 0.17.0
• datashader 0.12.0

Description of expected behavior and the observed behavior

I expect Canvas().raster to return a DataArray object, instead it returns an error.

Complete, minimal, self-contained example code that reproduces the issue

Adapted from your documentation

from datashader import transfer_functions as tf, reductions as rd
import numpy as np, datashader as ds, xarray as xr

def f(x,y):
    return np.cos((x**2+y**2)**2)

def sample(fn, n=50, range_=(0.0,2.4)):
    xs = ys = np.linspace(range_[0], range_[1], n)
    x,y = np.meshgrid(xs, ys)
    z   = fn(x,y)
    return xr.DataArray(z, coords=[('y',ys), ('x',xs)])

da = sample(f)
ds.Canvas().raster(da, interpolate='linear')

Stack traceback and/or browser JavaScript console output

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-3-7e8b8b432665> in <module>
     12 
     13 da = sample(f)
---> 14 ds.Canvas().raster(da, interpolate='linear')

/opt/conda/lib/python3.8/site-packages/datashader/core.py in raster(self, source, layer, upsample_method, downsample_method, nan_value, agg, interpolate, chunksize, max_mem)
   1131         dims = [ydim, xdim]
   1132         attrs = dict(res=res[0])
-> 1133         if source._file_obj is not None and hasattr(source._file_obj, 'nodata'):
   1134             attrs['nodata'] = source._file_obj.nodata
   1135 

/opt/conda/lib/python3.8/site-packages/xarray/core/common.py in __getattr__(self, name)
    237                 with suppress(KeyError):
    238                     return source[name]
--> 239         raise AttributeError(
    240             "{!r} object has no attribute {!r}".format(type(self).__name__, name)
    241         )

AttributeError: 'DataArray' object has no attribute '_file_obj'

I think this is related to #4809.

[jbednar]

Same issue as holoviz/hvplot#563 . Proposed fixes welcome; I don't know where to find nodata in an xarray object.

[michaelaye]

I'm afraid it depends on the data source.
In some of my data sources it looks like this:

This is the result of the 'open_rasterio" function, which uses gdal which is one of the most used geo-raster libs, so maybe that's quite reliable? However, that's only available because the array creator put it there. So, looking for it at obj.attrs['nodatavals'] would hit a lot of raster-data, but I'm not sure there's a standard attribute location for opening netCDF files for example?

[philippjfr]

This was always just a heuristic. I think we should support the major readers and conventions including attrs['NODATA'] and attrs['nodatavals'].

[ghost]

This was always just a heuristic. I think we should support the major readers and conventions including attrs['NODATA'] and attrs['nodatavals'].

attrs['nodata'] is another variation

[michaelaye]

So, lines 1133...:

if source._file_obj is not None and hasattr(source._file_obj, 'nodata'):
            attrs['nodata'] = source._file_obj.nodata

could maybe become something like:

nodata_keys = ['NODATA', 'nodatavals', 'nodata']
for key in nodata_keys:
    if hasattr(source.attrs, key):
        attrs['nodata'] = source.attrs[key]

?

[jbednar]

Yes, that's basically what I'm testing now.

[scottyhq]

FYI, this is how rioxarray currently does it for any given xarray object:

1. Look in attributes (attrs) of your data array for the _FillValue then missing_value then fill_value and finally nodata.
2. Look in the nodatavals attribute. This is for backwards compatibility with xarray.open_rasterio.

from:
https://corteva.github.io/rioxarray/stable/getting_started/nodata_management.html#Search-order-for-nodata-(DataArray-only):

[michaelaye]

Temporary workaround, if you just want the plot working without caring about nodata vals (maybe they can be set manually?):

arr['_file_obj'] = None

[michaelaye]

Temporary workaround, if you just want the plot working without caring about nodata vals (maybe they can be set manually?):

arr['_file_obj'] = None

[ppwadhwa]

This should be fixed in 0.12.1, which is available on the pyviz channel, conda-forge, and pip.