netCDF scalefactore is NaN resulting in an empty output variable about divand.jl

[Alexander-Barth]

I am able to reproduce the error in DIVAnd_climatology_10ya.jl that you are seeing. Surface data was with a depth of exactly zero was no more used. With my commit 2a5e733 this test now works. We will check other tests.
Thanks!

[Alexander-Barth]

ERROR: LoadError: KeyError: key NCDatasets.checksum not found

For your information, this issue is now also fixed in my branch.

[Alexander-Barth]

Just to get the obvious issue out of the way: how many observations do you have corresponding to sea points within your time range?

I just added some more diagnostic information to this function on my branch Alex, e.g.

]add DIVAnd#Alex 

(you can check with using DIVAnd; pathof(DIVAnd)).

Can you rerun your test-case with this branch and enabling debugging ENV["JULIA_DEBUG"] = "DIVAnd" ?

[Alexander-Barth]

These changes are now also on master.

[ax1ine]

Thank you, Alexander! I've enabled debugging and this is my output from the updated version:

nobs = 241105
length(obsval) = 241105
extrema(obslon) = (-179.9944, 179.9967)
extrema(obslat) = (50.0023, 87.985)
extrema(obstime) = (2016-01-01T00:56:00, 2016-01-30T22:51:00)
TS = TimeSelectorYearListMonthList{Array{UnitRange{Int64},1},Array{Array{Int64,1},1}}(UnitRange{Int64}[2016:2016],
[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]])
0.501255 seconds (520 allocations: 187.346 MiB, 11.08% gc time)
size(mask) = (719, 59, 2)
[ Info: Creating netCDF file
[ Info: Time step 1 / 1
┌ Debug: background_epsilon2_factor: 120552.5
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\diva.jl:460
┌ Debug: alphabc 0 for size (2,)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 1
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 0 or 0.0 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
┌ Debug: fbackground: [0.021922769747620075, 0.021922769747620075]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\diva.jl:481
[ Info: scaled correlation length (min,max) in dimension 1: (300000.0, 300000.0)
[ Info: scaled correlation length (min,max) in dimension 2: (300000.0, 300000.0)
[ Info: scaled correlation length (min,max) in dimension 3: (25.0, 25.0)
┌ Debug: ("cutter", [(10.79186970867863, 309.183177852104), (5.395922012953954, 5.395922012958835), (250.0, 250.0)], (719, 59, 2), [0.0, 0.0, 0.0], 32, :auto)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_cutter.jl:39
┌ Debug: biggestproblemdirectn [2.0e6 1.28e6 640000.0 1.25e6]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:59
┌ Debug: problemcut: 84842
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:69
┌ Debug: biggestproblemdirect: 640000.0
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:70
┌ Debug: csteps [0, 0, 0]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:83
┌ Debug: error method: cpme
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:109
[ Info: number of windows: 1
┌ Debug: window: 1, indices: ([1, 1, 1], [719, 59, 2], [1, 1, 1], [719, 59, 2], [1, 1, 1], [719, 59, 2])
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:120
┌ Warning: scalefactore has the value NaN
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_adaptedeps2.jl:72
[ Info: number of current observations: 241105
[ Info: number of ignored observations: 241105 (typically havigng a NaN value or outside of the domain)
[ Info: mean of diagR: 0.09999999999999985
[ Info: minimum of diagR: 0.1
2.264618 seconds (685.18 k allocations: 724.518 MiB, 4.40% gc time)

DIVAnd is indeed ignoring all the observations. Have there been any changes to the bathymetry/mask recently? My script had been working with no issues until I updated DIVAnd while trying to fix that latitude issue.

[Alexander-Barth]

I got your email. But it is possible to reduce your test case to a minimal one (e.g. using only 2 observations) and removing parts of the script unrelated to the error? This would help me a lot.

[ax1ine]

Alexander, maybe we could try the standard examples? They must've been tested many times so it'll be easier to trace if something is wrong with the code or with my installation. I tried to run 'DIVAnd_climatology_10ya.jl' and it also didn't work for me.
Here's the output:

[ Info: Loading data from 'big file' X:\DIVAnd\DIVAnd-example-data\BlackSea\Salinity.bigfile
size(value) = (139230,)
size(depthr) = (3,)
mean(lenz) = 11.555555555555548
[ Info: Output file: X:\DIVAnd\DIVAnd-example-data\BlackSea\output.nc
[ Info: Creating netCDF file
[ Info: Time step 1 / 8
┌ Debug: background_epsilon2_factor: 46966.666666666664
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\diva.jl:460
┌ Debug: alphabc 0 for size (3,)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 1
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 8432 or 59.84386089425124 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
┌ Debug: fbackground: [18.18804260637072, 18.296614334714732, 18.475035070066536]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\diva.jl:481
[ Info: scaled correlation length (min,max) in dimension 1: (200000.0, 200000.0)
[ Info: scaled correlation length (min,max) in dimension 2: (200000.0, 200000.0)
[ Info: scaled correlation length (min,max) in dimension 3: (10.0, 13.333333333333334)
┌ Warning: resolution (20.0) is too coarse for correlation length 10.0 in dimension 3 at indices CartesianIndex(4,
3, 1) (skipping further tests). It is recommended that the resolution is at least 2 times finer than the correlation length.
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\utils.jl:18
┌ Debug: ("cutter", [(11.739795676344466, 13.18656121498891), (8.993203354855417, 8.993203354940864), (0.4444444444444445, 0.5)], (76, 36, 3), [0.0, 0.0, 0.0], 3, :auto)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_cutter.jl:39
┌ Debug: biggestproblemdirectn [187500.0 120000.0 60000.0 117187.5]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:59
┌ Debug: problemcut: 8208
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:69
┌ Debug: biggestproblemdirect: 60000.0
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_fittocpu.jl:70
┌ Debug: csteps [0, 0, 0]
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:83
┌ Debug: error method: cpme
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:109
[ Info: number of windows: 1
┌ Debug: window: 1, indices: ([1, 1, 1], [76, 36, 3], [1, 1, 1], [76, 36, 3], [1, 1, 1], [76, 36, 3])
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:120
┌ Debug: call DIVAndrun
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndgo.jl:244
┌ Debug: alphabc 1.0 for size (76, 36, 3)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 3
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 11 or 0.30994646379261764 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
┌ Debug: alphabc 1.0 for size (76, 36, 3)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 3
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 11 or 0.30994646379261764 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
ERROR: LoadError: ArgumentError: reducing over an empty collection is not allowed

The tests 'DIVAnd_simple_example_4D.jl' and 'DIVAnd_simple_example_big3D.jl' (which apparently produce no output netCDF), worked without errors.

Tried 'DIVAnd_simple_example_big3D.jl' and got this:

┌ Debug: alphabc 1.0 for size (80, 50, 5, 12)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 2
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 387 or 19.35 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
9.840046 seconds (115.88 k allocations: 4.514 GiB, 7.44% gc time)
maxiter = 100
diagshift = 0.001
┌ Debug: alphabc 1.0 for size (80, 50, 5, 12)
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAndrun.jl:44
┌ Debug: effective number of dimensions (neff): 4
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\DIVAnd_background.jl:77
┌ Debug: number of point outside 387 or 19.35 %
└ @ DIVAnd C:\Users\ax_li.julia\packages\DIVAnd\S4iuk\src\sparse_interp.jl:139
14.683576 seconds (161.53 k allocations: 12.511 GiB, 14.97% gc time)
ERROR: LoadError: KeyError: key NCDatasets.checksum not found

Is there anything new with the NCDatasets package which hasn't been pushed to the master yet? I just one more time updated and rebuilt both DIVAnd and NCDatasets (0.10.2) - still the same errors... Could you, please, run those tests and confirm that they work correct with your installation?

[Alexander-Barth]

The tests 'DIVAnd_simple_example_4D.jl' and 'DIVAnd_simple_example_big3D.jl' (which apparently produce no output netCDF)

yes, this is correct. The result of the analysis is the returned to the user as output variable of the function DIVAndrun.

[Alexander-Barth]

For your information, all tests do not work again. Can you update to the latest version of my branch to see if it solves your issue too?

[ax1ine]

Yes, my issue is gone now, thank a lot, Alexander!