Open willtebbutt opened 3 weeks ago
Currently, there are many situation in which we run an entire sampler in order to test AD, such as many cases in https://github.com/TuringLang/Turing.jl/blob/master/test/mcmc/abstractmcmc.jl .
This is quite compute intensive. @yebai pointed out that perhaps we should just check that the gradient runs and is correct, rather than thousants of mcmc chains.
We should probably implement some functionality to benchmark various AD backends during this refactoring.
Currently, there are many situation in which we run an entire sampler in order to test AD, such as many cases in https://github.com/TuringLang/Turing.jl/blob/master/test/mcmc/abstractmcmc.jl .
This is quite compute intensive. @yebai pointed out that perhaps we should just check that the gradient runs and is correct, rather than thousants of mcmc chains.