Marginalapprox fix #6076
Marginalapprox fix #6076
Conversation
Codecov Report
@@ Coverage Diff @@
## main #6076 +/- ##
==========================================
+ Coverage 89.52% 89.53% +0.01%
==========================================
Files 72 72
Lines 12950 12948 -2
==========================================
Hits 11593 11593
+ Misses 1357 1355 -2
|
|
thanks @canyon289! |
pymc/gp/gp.py
Outdated
| **kwargs, | ||
| ) | ||
| approx_logp = self._build_marginal_likelihood_logp(y, X, Xu, noise, JITTER_DEFAULT) | ||
| pm.Potential(f"marginalapprox_logp_{name}", approx_logp) |
There was a problem hiding this comment.
This is the f"{name}_loglikelihood", right?
Just trying to find something more instructuve than "logp"
There was a problem hiding this comment.
good catch! yeah thanks it's not even correct because it is the log likelihood and not logp... fixed
pymc/tests/test_gp.py
Outdated
| self.gp = gp | ||
| self.sigma = 0.1 | ||
| self.x = np.linspace(-5, 5, 30) | ||
| self.y = 0.25 * self.x + self.sigma * np.random.randn(len(self.x)) |
There was a problem hiding this comment.
| self.y = 0.25 * self.x + self.sigma * np.random.randn(len(self.x)) | |
| self.y = np.random.normal(0.25 * self.x, self.sigma) |
|
It does address #5024, but I left it off since "ideally" you wouldn't use a potential here. If I could get something like DensityDist working, you would then include a |
This PR is meant to fix #5922 and possibly also #6069, but not 100% sure about the last one.
Ended up having to use
pm.Potentialinstead ofpm.DenisityDistto getpm.MarginalApprox.marginal_likelihoodworking properly. Usingpm.DensityDistit seems would involve introspecting the lazily evaluated covariance and mean functions for any random variables and their dims.I also tried to clean up the tests a bit for
MarginalApprox, because the reason #5922 took a while to be caught was becausefind_MAPor other any other methods weren't tested. #6069 does look to just be due to tolerances -- the results still look correct so I increased those a bit.Hopefully this PR can close #6069 and close #5922.