Incorrect (negative) values in safeRound result

[wiese]

For certain inputs (not entirely sure what they have in common) we are seeing results which do seem off and do not match Iteround's results for the same input.

Interestingly, a correct sum seems to be achieved by negative and positive values compensating for one another.

Please find a reproduction at https://github.com/wiese/saferound/commit/0bacbbc99c3853f7d28711e02b6f108c4c298092

The iteround equivalent (codesandbox here):

import iteround

data = {
    'a': 0.013245557004700679,
    'b': 0.7268604797100215,
    'c': 0.6155001637295456,
    'd': 0.06875329582046738,
    'e': 0.8387831576913838,
    'f': 0.03243429881206358,
    'g': 0.0023860988000495792,
    'h': 0.41960946536150784,
    'i': 0.675551119386058,
    'j': 0.6409523907667709,
    'k': 0.9659239729174308,
}

print(sum(data.values())) // 4.999999999999999

rounded = iteround.saferound(data, 0)

print(sum(rounded.values())) // 5.0
print(rounded.values()) // dict_values([0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0])

Please advise, am I missing something? Thank you!

[boredland]

Well, I guess I removed too much comparing with the original implementation. Pushed a bug fix that should align it. At least your test became green and while the order of things changed, the sum-result remained stable throughout the pre-exising tests.

[wiese]

Thanks for the swift reaction. The suggested regression test is now green.

It was my understanding that saferound tries to produce the same results as iteround.

However, after this change that's not the case for the pre-existing example in the test suite

input	4.0001, 3.2345, 3.2321, 6.4523, 5.3453, 7.3422
saferound	5, 3, 4, 6, 5, 7
iteround	4.0, 3.0, 3.0, 7.0, 6.0, 7.0

nor for the example of the iteround package

input	60.19012964572332, 15.428802458406679, 24.381067895870007
saferound	61, 15, 24
iteround	60.0, 16.0, 24.0

Is there a misunderstanding?

Thanks!

[boredland]

Yes, it was meant to be copy of the "DIFFERENCE" strategy of iteround. Perhaps I have overseen a sorting or something. Feel free to raise a PR ;-)

[boredland]

nvm, I think I have been able to fix it so, that the original basic difference test as well as the readme example are green.

[wiese]

Thanks for your work!

That last change addressed our concerns.