Cool. Do you use any well-known textbook? It would be best if we could test some well-designed hierarchical model in Anglican? I prefer the book Doing Bayesian Data Analysis, since it has some decently serious models there, yet is self-contained and approachable for users. I also have practically all other most popular textbooks, but they do not seem as well-balanced as this one. Even some really good books tend to use somewhat "easy" models that work well and fast while you are learning, even with not-so-fast tools, but once you try something more demanding - they are stuck. It is not even the problem with data points (I think 100 is not so little at all) because Bayesian tools are more useful the less data you have :) I am talking about model complexity due to hierarchy that tends to explode rather quickly...
On Sunday, October 23, 2016 at 10:13:48 PM UTC+2, Boris V. Schmid wrote: > > Not hierarchical, but continuous variables. It is our first foray into > bayesian inference, so we keep things somewhat simple. > > Can't give an exact comparison, but to run a model simulating a single > city (rats and fleas and human populations, no spatial component) is in the > order of minutes for my student working with PyMC, and fitting a mortality > curve based on ~100 datapoints. Myself, I was mostly playing along while > supervising, and that model in Anglican is stuck halfway an upgrade to use > clojure.as a testing framework. But as I recall, it also used to be in > the order of minutes. Will see if I can finish the upgrade and put it > online. > > On Sunday, October 23, 2016 at 8:45:47 PM UTC+2, Dragan Djuric wrote: > >> Are those hierarchical models? I also suppose the variables are >> continuous? What are typical running times for your analysis with Anglican, >> and what with PyMC? >> >> On Sunday, October 23, 2016 at 8:17:16 PM UTC+2, Boris V. Schmid wrote: >>> >>> I am using Anglican for estimating parameters of epidemiological models, >>> generally in the shape of limited (mortality) data, and less than a dozen >>> parameters that need to be simultaneously estimated. Works fine for that. A >>> good example of that type of problem is here: >>> http://www.smallperturbation.com/epidemic-with-real-data (but with >>> PyMC, a similar package for python). >>> >>> But you might be right that it won't hold in high-dimensional problems. >>> People in genomics are running models with many thousands of parameters >>> when trying to figure out how different genes contribute to a particular >>> cell phenotype. Don't think I would try that in Anglican :-). >>> >>> >>> On Sunday, October 23, 2016 at 6:06:49 PM UTC+2, Dragan Djuric wrote: >>>> >>>> Thanks. I know about Anglican, but it is not even in the same category, >>>> other than being Bayesian. Anglican also has MCMC, but, looking at the >>>> implementation, it seems it is useful only on smaller problems with >>>> straightforward and low-dimensional basic distributions, or discrete >>>> problems/distributions. I do not see how it can be used to solve even >>>> standard textbook examples in "real" bayesian data analysis. Otherwise, >>>> I'd >>>> use/improve Anglican, although its GPL license is a bit of a showstopper. >>>> >>>> I would loved to have been able to see how far Anglican can go >>>> performance-wise, and stretch it to its limits, though. However, it wasn't >>>> obvious how to construct any of more serious data analysis problems. >>>> Having >>>> seen its implementation, I expect the performance comparison would make >>>> Bayadera shine, so I hope I'll be able to construct some examples that can >>>> be implemented in both environments :) >>>> >>>> On Sunday, October 23, 2016 at 3:47:50 PM UTC+2, Boris V. Schmid wrote: >>>>> >>>>> Thanks Dragan. >>>>> >>>>> Interesting slides, and interesting section on Bayadera. Incanter, as >>>>> far as I know indeed doesn't support MCMC, but there is a fairly large >>>>> project based on clojure that does a lot of bayesian inference. >>>>> >>>>> Just in case you haven't run into it: >>>>> http://www.robots.ox.ac.uk/~fwood/anglican/examples/index.html >>>>> >>>>> (for the far future, there are some interesting developments happening >>>>> with approximate bayesian inference using neural network classification >>>>> to >>>>> speed things up. Fun stuff.) >>>>> >>>>> On Thursday, October 20, 2016 at 11:38:25 PM UTC+2, Dragan Djuric >>>>> wrote: >>>>>> >>>>>> Hi all, I posted slides for my upcoming EuroClojure talk, so you can >>>>>> enjoy the talk without having to take notes: >>>>>> http://dragan.rocks/articles/16/Clojure-is-not-afraid-of-the-GPU-slides-EuroClojure >>>>>> >>>>> -- You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com Note that posts from new members are moderated - please be patient with your first post. To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en --- You received this message because you are subscribed to the Google Groups "Clojure" group. To unsubscribe from this group and stop receiving emails from it, send an email to clojure+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
