Figure out discrepancy between different L likelihood+prior and difference

evidation-health / ContinuousTimeMarkovModel

Continuous-time Markov model with discrete observations

MIT License

23 stars 23 forks source link

Figure out discrepancy between different L likelihood+prior and difference #8

Closed pourzanj closed 9 years ago

pourzanj commented 9 years ago

When we print the logp of L along with it's prior and the logp+prior of the L proposal the difference is different than the difference that Theano spits out. If these differences are calculated incorrectly we could be accepting jumps in L that are too big. Jumps that are too big will lead to the sampling of X going awry since it is a complicated function of L.

pourzanj commented 9 years ago

Jumps were being accepted because the Jacobian of the inverse transform was being calculated incorrectly due to a Pymc3 bug. The issue has since been fixed.