[MNT] Fix test warnings in scripting part

[janezd]

Similar to #3502, this PR deals with most warnings in the scripting part, except those in postgresql and a few in the code I don't sufficiently understand.

[codecov]

Codecov Report

Merging #3503 into master will decrease coverage by 0.04%.
The diff coverage is 96.83%.

@@            Coverage Diff             @@
##           master    #3503      +/-   ##
==========================================
- Coverage   83.59%   83.54%   -0.05%     
==========================================
  Files         367      367              
  Lines       65580    65784     +204     
==========================================
+ Hits        54823    54961     +138     
- Misses      10757    10823      +66

[codecov]

Codecov Report

Merging #3503 into master will decrease coverage by <.01%.
The diff coverage is 94.73%.

@@            Coverage Diff             @@
##           master    #3503      +/-   ##
==========================================
- Coverage   83.97%   83.97%   -0.01%     
==========================================
  Files         370      370              
  Lines       66920    66941      +21     
==========================================
+ Hits        56196    56212      +16     
- Misses      10724    10729       +5

[janezd]

Build crashed (output too large) after removing the fix for np.sum(generator). Now I got to the bottom of it.

"once" filter in warnings has a nice undocumented behaviour: if the list of filters is changed, the registry of printed warnings is cleared. Hence any call to filterwarnings, even within catch_warnings context, resets the registry.

The problem in this PR is silencing the warnings in RowInstance.__init__ which resets the filters at every iteration of the loop over all instances in naive Bayesian classifier.

If you agree, I'll:

• I'll leave silencing of warnings within tests. It's harmless since the unittest framework resets the filters anyway.

• Almost all filters in actual code silence the scikit-sparse's use of matrix subclass. I would disable this warning for entire Orange, that is, put this filter to Orange.__init__. Benefits:

  • It avoids randomly resetting the filters.
  • We never want to see this warning anyway because it's never our fault and we can do nothing about it.
  • It will be more trivial to remove when it's fixed in scikit-sparse

• I will try to remove other filters, either by

  • fixing the code so they're not issued or
  • using np.errstate instead or
  • keeping filters, if they are unlikely to cause harm,

  or I may keep them, but hesitantly unless we expect them to go away some day.

When/If I get a green light from you, I'll first fix this PR and then do the same in #3502 (there seem to be just six places, also thanks to your cautious approach reviewing it).

[lanzagar]

At least now we know why filtering out some warnings can actually lead to more of them in the output!
I agree with all three points, and am a big fan of "fixing the code so they're not issued" approach anyway :)

[janezd]

I disabled warning about (scipy.sparse's) using of matrix for all Orange. Warnings are silenced on Travis and when tests are run locally in certain ways, but not always: if tests are run in PyCharm, warnings are re-enabled by the test runner. I don't see a good way of disabling it without hacking into the runner or warnings filter. This is not critical, though.

[janezd]

x[~mask] unnecessarily copies the array, and I wouldn't like to do it column by column. This patching is a non-idiomatic substitute for catch_warnings.

[janezd]

The warning here was "mean of empty slice", which appeared because median fell back to mean if size equalled 0 (achieving what?!)

[janezd]

I kept this catch_warnings although it's not in tests. It's supposed to be temporary.