Hello! I’m a newbie to Edward and Tensorflow so I decided to implement something very simple for a start. For my first exercise I am implementing bayesian inference for a Poisson distribution. This should be easy, right? Unfortunately I ran into problems using KLqp inference for this problem. I ca…

Hi! I am having the exact same problem. Have you solved it? @dustin has said a lot of times in this forum that ed.KLqp doesn’t work well with Gamma and Dirichlet variational approximations (eg: Error in inference.run() for a mixture model ), but I never found an explanation for this.

You can recover the posterior quite well if you use a log-normal approximation instead of a Gamma approximation: import numpy as np import tensorflow as tf import edward as ed from edward.models import Poisson, Gamma, Empirical ed.set_seed(42) # Generate sample data N = 100 lamb_true = 1.0 data =…

Problem with KLqp inference for Poisson distribution

pedrofale February 15, 2018, 2:57pm 4

Here’s a document about Gamma approximations in BBVI: ajbc.io/resources/bbvi_for_gammas.pdf

3 Likes

Topic		Replies	Views
Loss increasing in inference of Poisson Factorization	10	2256	February 21, 2018
A toy normal model failed (klqp) and why?	2	1704	July 25, 2017
Inference in mixture models containing transformations	2	1209	January 11, 2018
Simple Hierarchical Distribution	2	1549	October 12, 2017
KLqp with Beta/Bernoulli (Coin Flip example)	1	979	September 13, 2018

Problem with KLqp inference for Poisson distribution

Related topics