Hi Dustin and members of the Edward forum, I have been able to implement a zero-inflated Poisson model using variational inference. I replaced the Bernoulli distribution with the relaxed Bernoulli distribution, which is a continuous probability density function. I’ve inspected the parameter updates…

[image] mac449: Gamma Replace Gamma with Log-Normal Distribution and let’s check the new the results.

Using RelaxedBernoulli for a Zero-Inflated Poisson model

mac449 February 25, 2018, 7:56pm 3

Hi Ermia,

The Relaxed Bernoulli is unstable even with the lognormal distribution. The Bernoulli using MAP performs better.

Best,
Mark

1 Like

Topic		Replies	Views
Bernoulli distribution parameterized by a mapped Gaussian	11	1848	February 11, 2018
Variational Inference with Composition of Variables	6	1155	March 8, 2018
Implementation of ZIP model in Edwards	8	2308	April 18, 2018
Poisson-gamma regression	6	2033	March 19, 2017
Basic Regression Not Working	6	882	August 15, 2017