Problem with boundary values in .core.grids._horizontal_gradient

[lesommer]

Problem with boundary values in .core.grids._horizontal_gradient

[kuchaale]

@lesommer : I understand your issue with the specification of np.nan. However, when I run the code below (without dtype specification):

`import numpy as np
import xarray as xr

def is_numpy(array):
test = bool( isinstance(array,np.ndarray)
+ isinstance(array,np.ma.masked_array) )
return test

def _horizontal_gradient(scalararray):
data = scalararray.data
coords = scalararray.coords
dims = scalararray.dims
chunks = scalararray.chunks
if is_numpy(data):
da_dj,da_di = np.gradient(data)
else:
x_derivative = lambda arr:np.gradient(arr,axis=-1) # req. numpy > 1.11
y_derivative = lambda arr:np.gradient(arr,axis=-2) # req. numpy > 1.11
gx = data.map_overlap(x_derivative,depth=(0,1),boundary={1: np.nan})#, dtype=data.dtype)
gy = data.map_overlap(y_derivative,depth=(1,0),boundary={0: np.nan})#, dtype=data.dtype)
da_di = xr.DataArray(gx,coords,dims)
da_dj = xr.DataArray(gy,coords,dims)
return da_dj,da_di

ds = xr.open_dataset('../data/heigh_2017011715Z_1000hPa.nc')
ds = ds.chunk(chunks = 40)
ds`

I get this error:
ValueError: dtypeinference failed inmap_blocks`.

Original error is below:

ValueError('Shape of array too small to calculate a numerical gradient, at least two elements are required.',)

Traceback:

File "c:\python27\lib\site-packages\dask\array\core.py", line 457, in apply_infer_dtype
o = func(*args, **kwargs)
File "", line 26, in
x_derivative = lambda arr:np.gradient(arr,axis=-1) # req. numpy > 1.11
File "c:\python27\lib\site-packages\numpy\lib\function_base.py", line 1642, in gradient
"Shape of array too small to calculate a numerical gradient, "
`
This issue is induced by the non-specification of dtype.

[kuchaale]

I think your current version would not work anyway if you don't specify a dtype explicitly, see my issue here.

[lesommer]

@kuchaale : thanks for your comment and your suggestion. But I am not sure to get what you mean by "would not work" here. The present version of the code actually produce a result, but does not compute the gradient close to the boundary (thanks to the specification of np.nan).

[redouanelg]

Hi @kuchaale, your solution worked for me, please open a PR or allow me to do it