search

torch / torch7

http://torch.ch
Other
9k stars 2.38k forks source link

QR tests are failing because of tolerance #376

Open bartvm opened 9 years ago

bartvm commented 9 years ago

On two separate machines (but both using OpenBLAS) some the QR tests fail when using single precision, all falling short of the given tolerance by a small margin: https://gist.github.com/bartvm/90b11759edbfe23d796f

soumith commented 9 years ago

@bartvm should I just bump up the tolerance?

bartvm commented 9 years ago

Yeah, I guess that the tolerance is just slightly too high for some BLAS libraries. Looks like 2e-05 would do the trick already.

andreaskoepf commented 8 years ago

Shouldn't we have deterministic unit tests (I think @hughperkins already mentioned this)? This could be achieved by explicitly setting the rng seed value at the beginning of each unit test. If we want to do some extensive testing to include some unlikely constellations we could create a procedure that runs each test 100 or 1000 times with different (but still 100% reproducible) seed-values.