Review approach to fixing rng state for each test case

pyro-ppl / brmp

Bayesian Regression Models in Pyro

Apache License 2.0

70 stars 8 forks source link

Review approach to fixing rng state for each test case #16

Closed null-a closed 5 years ago

null-a commented 5 years ago

I copied the code for this verbatim from pyro in order get the tests working standalone.

Tests were flakier than expected without this
Assuming we continue to fix the seed, is there anything to do for numpyro?

neerajprad commented 5 years ago

is there anything to do for numpyro?

In NumPyro, we explicitly need to pass an rng key to inference algorithms and in fact to all distribution .sample methods, so it should be fully deterministic, unless we are using Python's random state to sample this value from.