Skill groupby attrs

[jsmariegaard]

Extracted by="attrs:gtype" part of #331

e.g. "attrs:gtype" or "attrs:DA" (to distinguish between assimilation and validation stations)

[jsmariegaard]

The to_dataframe() method returned object dtypes which I have now changed to category. This however leads to problems with the groupby specifically for cc.mean_skill() in multiple variable cases. I tried different things including setting observed=True, but that gives problems in gridded_skill() (where empty bins makes sense). I think that the actual problem is that the "default" by that mean_skill() passed on to skill() is ["model", "observation", "variable"] even though an observation can only have one variable. So I guess the by should instead be ["model", "observation"] and then just add variable afterwards. Maybe we should even check that observation and variable do not both occur in the by?

[ecomodeller]

Very useful functionality! 👍

Here is a snippet of a slightly incomplete example, where I have added attrs to 2(3) observations.

When the attrs is absent, the default is now to exclude it from the skill table, but by setting observed=True

...
>>> o1 = ms.PointObservation('HKNA_Hm0.dfs0', attrs={"use": "calibration"})
>>> o2 = ms.PointObservation("eur_Hm0.dfs0",   attrs={"use": "validation"})
>>> o3 = ms.TrackObservation("Alti_c2_Dutch.dfs0")
>>> cc = ms.match(obs=[o1, o2, o3], mod=[mr1, mr2])
>>> cc.skill(by=("model","attrs:use"), observed=False).round(2)
                     n  bias  rmse  urmse   mae    cc    si    r2
model use
SW_1  False        113 -0.00  0.35   0.35  0.29  0.97  0.13  0.90
      calibration  386 -0.19  0.35   0.29  0.25  0.97  0.09  0.91
      validation    67 -0.07  0.22   0.21  0.19  0.97  0.08  0.93
SW_2  False        113  0.08  0.43   0.42  0.36  0.97  0.15  0.85
      calibration  386 -0.10  0.29   0.28  0.21  0.97  0.09  0.93
      validation    67 -0.00  0.23   0.23  0.20  0.97  0.09  0.93
>>> cc.skill(by=("model","attrs:use")).round(2)
                     n  bias  rmse  urmse   mae    cc    si    r2
model use
SW_1  calibration  386 -0.19  0.35   0.29  0.25  0.97  0.09  0.91
      validation    67 -0.07  0.22   0.21  0.19  0.97  0.08  0.93
SW_2  calibration  386 -0.10  0.29   0.28  0.21  0.97  0.09  0.93
      validation    67 -0.00  0.23   0.23  0.20  0.97  0.09  0.93

[jsmariegaard]

Finally managed to do the merge 😬that was not easy

[ecomodeller]

It seems overly specific to assert that the index is sorted in this order.

Isn't it enough to verify that:

assert "point" in sk.index
assert "track" in sk.index

[jsmariegaard]

true - will fix in next PR which will be on sorting