Variational inference engines

Question

After doing some research on the topic, I have noticed a surprising deficit of inference packages and libraries that rely on message-passing or optimization methods for Python and R.

To the best of my knowledge, these methods are extremely useful. For example, for a Bayes Network (directed, acyclic) belief-propagation alone should be able to give exact answers. However, most inference software that is available online (e.g. STAN, BUGS, PyMC) rely on MCMC methods.

In the Python case, to the best of my knowledge, neither PyMC, scikit-learn or statsmodels include variational inference algorithms such as belief propagation, message-passing methods or any of their variants.

Why is that? Are these methods less used in practice because they are seen not as powerful or generic as their MCMC counterparts? or Is it simply a matter of lack of manpower and time?

Answer 1

Have you looked at Edward? The Inference API supports among other things Variational inference:

• Black box variational inference
• Stochastic variational inference
• Variational auto-encoders
• Inclusive KL divergence: KL(p∥q)

Answer 2

Some years have passed. STAN now implements ADVI (gradients of the ELBO with reparameterization trick), using the vb command (which I guess stands for Variational Bayes). E.g., in R:

library("rstan") # load the rstan library
fit = vb(model, data, output_samples = 20000, adapt_iter = 10000 ,init = list('param1' = param1, ...), seed)

Pyro is a python library that implements BBVI (gradients of the ELBO with log-derivative trick + variance reduction techniques). Main code is:

import pyro
auto_guide = pyro.infer.autoguide.AutoNormal(model)
adam = pyro.optim.Adam({"lr": 0.02})
elbo = pyro.infer.Trace_ELBO()
svi = pyro.infer.SVI(model, auto_guide, adam, elbo)
for step in range(1000): 
    loss = svi.step(data)