In my quest to learn more about Bayesian Analysis, I spent the last few weeks working through the book Bayesian Analysis with Python - 2nd Edition by Osvaldo Martin. In their great wisdom, Amazon now only allows reviews by people who have been spending "real" money on Amazon, as opposed to those like me using "fake" money in the form of gift cards and credit card rewards. So I guess I will have to hold off posting my review of this great book until my fake money runs out. Meanwhile, you can read other people's review on Amazon. In this post, I would like to share my notebooks where I have worked out the exercises for Chapters 2-7 (there are 9 chapters, the first one is an introductory one, and the last two chapters cover areas only indirectly related to Bayesian Analysis, or more specifically, Bayesian analysis using Monte Carlo Markov Chain (MCMC) techniques. Here are my notebooks on Github. No guarantees, of course, since I am learning as well.
The author Osvaldo Martin is one of the developers of the ArviZ project, an exploratory analysis tool for Bayesian models. The ArviZ project was spun-off from the PyMC3 project, and many PyMC3 calls such as pm.traceplot() are actually calls to az.plot_trace() under the hood. According to the ArviZ website, it also supplies functionality for other Bayesian libraries, such as PyStan, Pyro, and TF Probability. Martin is also a PyMC3 maintainer, and the book reflects that. Compared with my previous foray into PyMC3, where I was trying to convert JAGS models taught in the two Coursera courses from the University of California, Santa Cruz (UCSC), into equivalent PyMC3 models using online documentations and other Internet resources, this time the focus was on exploring various code features offered by PyMC3 and ArviZ.
The book is billed as an introductory one on the PyMC3 site, but I thought it had good followup material for someone who has already taken the two UCSC Coursera courses. The focus on those courses was on the theory, to explain why you would choose one distribution over the other, linear and logistic regression, and some exposure to hierarchical and mixture models. This book contains a more code-heavy exposition of all these subjects, plus an introduction to Dirichlet processes as part of the Mixture models chapter, and an introduction to Gaussian processes.
Of course, I still have much more to learn about all this. Just today, I came across a link to Nicole Carlson's PyMC3 talk at PyData Chicago 2016, where she walks through some toy problems with PyMC2. She introduces the ideas of using Theano shared variables for input and output here, so you can use the model in a more Scikit-Learn like way, training with one dataset and evaluating with another. She also covers the newer PyMC3 Variational Inference API (ADVI), an alternative to its older MCMC API, and how to use the output of the ADVI model as input to an MCMC sampling model. In addition, she also describes serializing a PyMC3 model by pickling the trace and using it as a generative model. I also found the notebooks for her subsequent talk at PyData Chicago 2017, where she describes her work to wrap PyMC3 models in a more familiar Scikit-Learn like interface.
Another recommendation from the PyMC3 site, as well as from this book, was the "puppy book" Doing Bayesian Data Analysis by John Kruschke. Fortunately, that is available in my employer's online library, so I guess that is another book I won't buy from Amazon with fake or real money.
Thats pretty much all I had for today. Obviously the best way to learn this stuff is to try doing it yourself, but sharing these notebooks on Github in case someone gets stuck and can benefit from them. In case you do end up using them, please let me know what you think.
No comments:
Post a Comment
Comments are moderated to prevent spam.