meta: Testing probabilistic programs
rlouf opened this issue · 2 comments
rlouf commented
It is not clear how we should implement tests for probabilistic models due to the randomness associated with the sampling process. Currently, the tests check that the API works as expected, but not the validity of the outputs.
The following blog article has some interesting ideas:
https://alexey.radul.name/ideas/2016/on-testing-probabilistic-programs/
twiecki commented
In PyMC3 we have regression tests that check numerical output and set the seed.
rlouf commented
That makes sense, thanks for the pointer.