Closed carlini closed 3 years ago
Would setting a fix seed per test address be a satisfactory solution?
Does jax guarantee that the same seed will generate the same random sequence across minor versions of the library? If so, then maybe.
The seed should be reset before every random call though, or else changing one test could make a new one fail.
Well we can pass a generator per test so the seed can be controlled locally without trouble. As to JAX guarantees, I don't know.
I think we can close this now. @david-berthelot
There are currently tests in tests/testrandom.py that fail with small probability (and some with not-so-small probability)
For example, here
https://github.com/google/objax/blob/07b2bb09980ee92935973015ea10d1fa99e5e766/tests/testrandom.py#L54-L55
the probability that this fails is ~11% (taking the mean of 1000*100 samples with man stdev=2 should be distributed like a normal with mean 006.324e-3 and this is greater than .01 with probability 11%).
The other tests also fail with non-negligible probability.