Getting scalar values of the weighted average when input is 2D and weights is a 2D array about esmlab

[mnlevy1981]

A better version of the script:

#!/usr/bin/env python

import xarray as xr
import numpy as np
import esmlab

# Read in WOA observational data (in 12 files)
in_files = []
for n in range(0,12):
  in_files.append("/glade/work/mclong/woa2013v2/POP_gx1v7/woa13_all_n{:02d}_gx1v7.nc".format(n))
ds = xr.open_mfdataset(in_files, decode_times=False)

# Reduce dataset
# 1. Want NO3 averaged over all files
field = ds.NO3.isel(z_t=0).mean('time')
# 2. TAREA is identical across files, so just use the first time dimension of it
TAREA = ds.TAREA.isel(time=0)

# Statistics
print("field min: {:.2f}".format(np.nanmin(field)))
print("field max: {:.2f}".format(np.nanmax(field)))
print("field weighted mean:\n{}".format(esmlab.statistics.weighted_mean(field, TAREA)))

Output:

/glade/work/mlevy/miniconda3/envs/NPL-conda/lib/python2.7/site-packages/dask/array/numpy_compat.py:28: RuntimeWarning: invalid value encountered in divide
  x = np.divide(x1, x2, out)
field min: 0.00
field max: 33.95
field weighted mean:
<xarray.DataArray ()>
dask.array<shape=(), dtype=float64, chunksize=()>
Coordinates:
    z_t      float64 500.0
    time     float64 6.0

(I don't think the numpy_compat.py warning affects anything... numpy doesn't complain about it with nanmin and nanmax, and I can still plot field in other scripts)

[andersy005]

@mnlevy1981, to make sure I am getting this, what should be the expected output's dimensions (shape) of esmlab.statistics.weighted_mean(field, TAREA)?

With

# 1. Want NO3 averaged over all files
field = ds.NO3.isel(z_t=0).mean('time')
# 2. TAREA is identical across files, so just use the first time dimension of it
TAREA = ds.TAREA.isel(time=0)

field, and TAREA are as follows:

In [20]: field.shape
Out[20]: (384, 320)

In [21]: TAREA.shape
Out[21]: (384, 320)

[mnlevy1981]

It should be a scalar quantity. I just realized I could come up with an even simpler example using 1D arrays rather than reading data from disk, but I need to run... I'll try to find time tonight to post more. But the gist is that if

TAREA = [1, 1, 2]
field = [1, 3, 8]

Then I expect the weighted mean to be 5: the weighted sum of field is 20, and the weights sum to 4 so the weighted mean is 20/4 = 5.

[andersy005]

I think I figured it out:

1. Due to xr.open_mfdataset(), xarray loads data into dask arrays instead of in memory NumPy arrays. The implication of this is that any subsequent operation is likely to return an xarray object with dask arrays.
2. There are two solutions to this:
   • Force xarray to load the entire data set in memory with

ds = xr.open_mfdataset(in_files, decode_times=False).load()

Or

• Tell Dask to compute and return and an in-memory result

w_mean = esmlab.statistics.weighted_mean(field, TAREA).data.compute()

In [47]: field
Out[47]:
<xarray.DataArray 'NO3' (nlat: 384, nlon: 320)>
array([[      nan,       nan,       nan, ...,       nan,       nan,       nan],
       [      nan,       nan,       nan, ...,       nan,       nan,       nan],
       [22.465448, 22.383131, 22.292166, ...,       nan,       nan,       nan],
       ...,
       [      nan,       nan,       nan, ...,       nan,       nan,       nan],
       [      nan,       nan,       nan, ...,       nan,       nan,       nan],
       [      nan,       nan,       nan, ...,       nan,       nan,       nan]],
      dtype=float32)
Coordinates:
    TLAT     (nlat, nlon) float64 -79.22 -79.22 -79.22 ... 72.2 72.19 72.19
    TLONG    (nlat, nlon) float64 -39.44 -38.31 -37.19 ... -41.08 -40.65 -40.22
    z_t      float64 500.0
Dimensions without coordinates: nlat, nlon

In [48]: TAREA
Out[48]:
<xarray.DataArray 'TAREA' (nlat: 384, nlon: 320)>
dask.array<shape=(384, 320), dtype=float64, chunksize=(384, 320)>
Coordinates:
    TLAT     (nlat, nlon) float64 -79.22 -79.22 -79.22 ... 72.2 72.19 72.19
    TLONG    (nlat, nlon) float64 -39.44 -38.31 -37.19 ... -41.08 -40.65 -40.22
    time     float64 6.0
Dimensions without coordinates: nlat, nlon
Attributes:
    long_name:  area of T cells
    units:      cm^2

In [49]: w_mean = esmlab.statistics.weighted_mean(field, TAREA)

In [51]: w_mean
Out[51]:
<xarray.DataArray ()>
dask.array<shape=(), dtype=float64, chunksize=()>
Coordinates:
    z_t      float64 500.0
    time     float64 6.0

In [52]: w_mean.data
Out[52]: dask.array<truediv, shape=(), dtype=float64, chunksize=()>

In [53]: w_mean.data.compute()
Out[53]: 5.146317884602825

In [54]: np.nanmean(field)
Out[54]: 5.7132335

[andersy005]

@mnlevy1981,
Oooh one more thing, if you want your result to be an xarray object,

do:

w_mean = esmlab.statistics.weighted_mean(field, TAREA).load()

instead of:

w_mean = esmlab.statistics.weighted_mean(field, TAREA).data.compute()

[mnlevy1981]

Thanks! The line

print("field weighted mean: {:.2f}".format(esmlab.statistics.weighted_mean(field, TAREA).load().values))

Works as expected. Is accessing values directly the preferred method of getting a scalar out of an xarray object? It looks like

print("field weighted mean: {:.2f}".format(esmlab.statistics.weighted_mean(field, TAREA).load().item()))

is another option.

[andersy005]

I think this is the appropriate way:

print("field weighted mean: {:.2f}".format(esmlab.statistics.weighted_mean(field, TAREA).load().values))

xref: http://xarray.pydata.org/en/stable/generated/xarray.DataArray.values.html#xarray.DataArray.values

[andersy005]

@mnlevy1981, should I close this? Or free feel to close it if you are satisfied with the conclusion from our last week's meeting.