mjhajharia / transforms

2 stars 1 forks source link

init values for sampling #60

Open mjhajharia opened 1 year ago

mjhajharia commented 1 year ago

so for the hyperspherical.stan and hypersphericalangular.stan default initialization fails and something very close to 0 works.

now the question being, would it make sense to use a modified init that would work for all the transforms and report that in the standard results and the other modifications that we try in the separate discussion for results?

or does it make sense to check the init correlation with plots for every transform and imperfectly choose the best initialization for each transform and try to justify why.

cc: @sethaxen @bob-carpenter @spinkney

bob-carpenter commented 1 year ago

Rather than making special adjustments for transforms, I think we want to discuss how robust they are to default initializations. If they don't work well with default initializations, they won't work in the tails of distributions.