Fix of gauss_2d_large(seed=63) -> NaN

[abukaj]

kcsd.validation.csd_profile.gauss_2d_large(seed=63) does not return
NaN anymore.

repeatUntilValid() decorator has been defined for that purpose.

A simple test for the fix provided in the __main__ section of the module.

Note: The issue has not been solved by fixing distribution of
the zs variable in order to provide backward-compatibility.

[coveralls]

Pull Request Test Coverage Report for Build 159

• 18 of 25 (72.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.09%) to 60.271%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
kcsd/validation/csd_profile.py	18	25	72.0%

Totals
Change from base Build 154:	0.09%
Covered Lines:	1693
Relevant Lines:	2809

---

💛 - Coveralls

[ccluri]

Nice workaround!
This should apply to 1D and 3D, small and large as well - right?

Although I only tested for seeds between 1-100, maybe the seeds else where have the same issue?

[abukaj]

@ccluri Thanks! I have not checked the other functions but I will do in a moment. A respective commit shall arrive today.

[ccluri]

I would say use something like
seed = rstate.randint(2**32 - seed)
Just so its not the same seed for all the failed candidates.
Over engineering - yes.

[abukaj]

@ccluri
The 2**32 argument of rstate.randint() is the upper bound (exclusive) on the random int. Also on the seed parameter of np.random.RandomState constructor.

The actual seed for rstate.randint() is given when the rstate object is created. The idea is to have a 'seedchain' seed[n+1] = randint(2**32, seed=seed[n]).

BTW: I am testing the csd_profile functions with the following code (I hope):

import numpy as np
from kcsd import csd_profile as CSD
import collections
x1d = np.mgrid[0.:1.:10j]
x2d = np.mgrid[0.:1.:10j, 0.:1.:10j]
x3d = np.mgrid[0.:1.:10j, 0.:1.:10j, 0.:1.:10j]
seed = 0
fails = []
fs = [CSD.gauss_1d_mono, CSD.gauss_1d_dipole, CSD.gauss_2d_large, CSD.gauss_2d_small, CSD.gauss_3d_small, CSD.gauss_3d_large]
while True: 
     if seed % 10000 == 0: 
         print(f'Fails for seed < {seed}:') 
         for name, n in collections.Counter(a for a,_ in fails).items(): 
             print(f' {name} : {n}') 
              
     for f in fs: 
         try: 
             res = f(x3d, seed) 
         except ValueError: 
             try: 
                 res = f(x2d, seed) 
             except ValueError: 
                 res = f(x1d, seed) 
         if not np.isfinite(res).any(): 
             fails.append((f.__name__, seed)) 
             #print(f"{f.__name__}({seed})") 
              
     seed += 1 

So far (for 0 <= seed < 600000) only gauss_2d_large() has failed (1158 times).

[ccluri]

Okay, let me know when you are content with the update - I am happy to merge when you are.

[abukaj]

@ccluri I am happy with gauss_2d_large() implementation in f59c426. Did you say something about over engineering? ;)
But let's wait till tomorrow for results for other non-decorated functions. So far seeds < 4840000 caused no problem.

[abukaj]

@ccluri I have tested 46600000 seeds. No NaNs returned by other functions.