Closed AlexAndorra closed 1 month ago
Check out this pull request on
See visual diffs & provide feedback on Jupyter Notebooks.
Powered by ReviewNB
@bwengals can we merge this one please?
View / edit / reply to this conversation on ReviewNB
bwengals commented on 2024-05-27T05:04:43Z ----------------------------------------------------------------
The truncated normal makes sense to me for the lengthscales, but it seems a bit strange for eta and sigma, because it won't let either the GP or the noise go to zero. I think it'd make more sense to use HalfNormals or something like that for those parameters. Does it still sample with those priors?
AlexAndorra commented on 2024-05-27T18:25:51Z ----------------------------------------------------------------
Yep, it still samples with pm.HalfNormal("sigma", 0.5)
and pm.HalfNormal("eta", 0.5)
But eta and sigma can't be equal to zero anyways, can they? What I'm trying to do is avoid the near-zero region. Does that make sense?
bwengals commented on 2024-05-27T22:32:07Z ----------------------------------------------------------------
I think its analogous to doing a linear regression, and putting a Normal(0, sigma) prior on the beta coefficients, and a half-normal or something on the likelihood sigma, right? Or Im missing something.
Like it'd be strange to do y ~ N(beta0 + beta1 * x, sigma)
with truncated normal priors on beta0
, beta1
and sigma
? Unless you had a very particular reason
AlexAndorra commented on 2024-05-28T13:40:24Z ----------------------------------------------------------------
Yep, agreed. I've definitely done the second case -- it's very useful when you have prior information about the parameters. Here though it's hard to argue for or against, as it's a simulated case.
Anyways, that's not a blocker and runs fine now, so we can go ahead
Ah sorry for missing this one, left one quick Q then all good.
Yep, it still samples with pm.HalfNormal("sigma", 0.5)
and pm.HalfNormal("eta", 0.5)
But eta and sigma can't be equal to zero anyways, can they? What I'm trying to do is avoid the near-zero region. Does that make sense?
View entire conversation on ReviewNB
Thanks @bwengals ! Just pushed the changes. I also had a small question on your comment, but this is not a blocker, so feel free to merge
I think its analogous to doing a linear regression, and putting a Normal(0, sigma) prior on the beta coefficients, and a half-normal or something on the likelihood sigma, right? Or Im missing something.
Like it'd be strange to do y ~ N(beta0 + beta1 * x, sigma)
with truncated normal priors on beta0
, beta1
and sigma
? Unless you had a very particular reason
View entire conversation on ReviewNB
Can you approve and merge @bwengals if that looks good to you?
Yep, agreed. I've definitely done the second case -- it's very useful when you have prior information about the parameters. Here though it's hard to argue for or against, as it's a simulated case.
Anyways, that's not a blocker and runs fine now, so we can go ahead
View entire conversation on ReviewNB
Various updates and improvements on the Kronecker variance GP example:
📚 Documentation preview 📚: https://pymc-examples--653.org.readthedocs.build/en/653/