https://dfm.io/ Recent content on Dan Foreman-Mackey Hugo -- gohugo.io en-us © 2014-{year} Dan Foreman-Mackey Thu, 28 Jul 2022 00:00:00 +0000 https://dfm.io/posts/intro-to-numpyro/ Thu, 28 Jul 2022 00:00:00 +0000 https://dfm.io/posts/intro-to-numpyro/ Over the past year or so, I’ve been using JAX extensively for my research, and I’ve also been encouraging other astronomers to give it a try. In particular, I’ve been using JAX as the computation engine for probabilistic inference tasks. There’s more to it, but one way that I like to think about JAX is as NumPy with just-in-time compilation and automatic differentiation. The just-in-time compilation features of JAX can be used to speed up you NumPy computations by removing some Python overhead and by executing it on your GPU. https://dfm.io/posts/simple-python-module/ Thu, 18 Mar 2021 00:00:00 +0000 https://dfm.io/posts/simple-python-module/ This post is an introduction to packaging Python code aimed at scientists (although the advice is probably more general) who want to be able to reuse Python functions that they have written as part of a common Jupyter notebook-based development workflow. It is not meant as a complete discussion of Python packaging for scientific software. Some day I’d love to write more about that (because I have so many thoughts and opinions), but in the meantime I’ll direct you to other resources like the excellent OpenAstronomy Python packaging guide, if you want more details. https://dfm.io/posts/extending-jax/ Mon, 11 Jan 2021 00:00:00 +0000 https://dfm.io/posts/extending-jax/ The primary version of this post can be found on GitHub: dfm/extending-jax This repository is meant as a tutorial demonstrating the infrastructure required to provide custom ops in JAX when you have an existing implementation in C++ and, optionally, CUDA. I originally wanted to write this as a blog post, but there’s enough boilerplate code that I ended up deciding that it made more sense to just share it as a repo with the tutorial in the README, so here we are! https://dfm.io/posts/pymc-pytorch/ Tue, 15 Oct 2019 00:00:00 +0000 https://dfm.io/posts/pymc-pytorch/ This post is a small extension to my previous post where I demonstrated that it was possible to combine TensorFlow with PyMC3 to take advantage of the modeling capabilities of TensorFlow while still using the powerful inference engine provided by PyMC3. The basic procedure involved writing a custom Theano operation that understood how to evaluate a TensorFlow tensor. In this post, I provide a similar snippet that can be used to combine PyTorch and PyMC3 to a similar end. https://dfm.io/posts/pymc3-mass-matrix/ Thu, 11 Oct 2018 00:00:00 +0000 https://dfm.io/posts/pymc3-mass-matrix/ In my work I often come across probabilistic models where there are strong correlations between parameters. This certainly isn’t a special feature of the problems that I work on, and the general advice from MCMC practitioners is that we should reparameterize our models in some form that removes the covariances, but this isn’t always practical. One of the reasons why emcee has been popular in astrophysics (I think) is that it uses an “affine invariant” algorithm. https://dfm.io/posts/emcee-pymc3/ Tue, 21 Aug 2018 00:00:00 +0000 https://dfm.io/posts/emcee-pymc3/ In this post, I will demonstrate how you can use emcee to sample models defined using PyMC3. Thomas Wiecki wrote about how to do this this with an earlier version of PyMC, but I needed an update since I wanted to do a comparison and PyMC’s interface has changed a lot since he wrote his post. This isn’t necessarily something that you’ll want to do (and I definitely don’t recommend it in general), but I figured that I would post it here for posterity. https://dfm.io/posts/pymc-tensorflow/ Thu, 02 Aug 2018 00:00:00 +0000 https://dfm.io/posts/pymc-tensorflow/ the most ambitious crossover event in history https://dfm.io/posts/autocorr/ Mon, 16 Oct 2017 00:00:00 +0000 https://dfm.io/posts/autocorr/ This is a cross post from the new emcee documentation. In this tutorial, we will discuss a method for convincing yourself that your chains are sufficiently converged. This can be a difficult subject to discuss because it isn’t formally possible to guarantee convergence for any but the simplest models, and therefore any argument that you make will be circular and heuristic. However, some discussion of autocorrelation analysis is (or should be! https://dfm.io/posts/travis-latex/ Wed, 12 Jul 2017 00:00:00 +0000 https://dfm.io/posts/travis-latex/ keeping an up-to-date build of your TeX source using GitHub and Travis https://dfm.io/posts/fitting-a-plane/ Thu, 01 Jun 2017 00:00:00 +0000 https://dfm.io/posts/fitting-a-plane/ Who knew that it was possible to say more about how to fit a line to data? Recently, Megan Bedell asked about how to generalize the method described in section 7 of the bible for fitting a line to data with 2-dimensional error elipses to higher dimensional problems, and how to include intrinsic scatter in that relation. While the correct generalization might be obvious to some of you, I wanted to go through the full derivation to make sure that I knew what was going on. https://dfm.io/posts/stan-c++/ Thu, 04 May 2017 00:00:00 +0000 https://dfm.io/posts/stan-c++/ The key feature that was missing from PyStan https://dfm.io/posts/exopop/ Mon, 06 Jul 2015 00:00:00 +0000 https://dfm.io/posts/exopop/ Warning this is an extremely technical blog post (even for me) but I think that it’s an interesting experiment so I thought I would post it anyways! The Kepler Mission has been a great success story for open science. The recent data releases have been on a fast and regular schedule and the data products are sometimes made public before the science office publishes their papers. Therefore, when the most recent installment in the “rate of Earths” saga—a discussion close to my heart—hit the ArXiv as Burke et al. https://dfm.io/posts/aas-hack-day/ Sun, 18 Jan 2015 00:00:00 +0000 https://dfm.io/posts/aas-hack-day/ The wrap-up post from the hack day at the 225th AAS meeting https://dfm.io/posts/mixture-models/ Fri, 12 Dec 2014 00:00:00 +0000 https://dfm.io/posts/mixture-models/ This post can be cited with the DOI: 10.5281/zenodo.15856 There are a lot of reasons why you might use a mixture model and there is a huge related literature. That being said, there are a few questions that I regularly get so I thought that I would write up the answers. In astronomy, the most common reason for using a mixture model is to fit data with outliers so that’s the language I’ll use but the results are applicable to any other mixture model. https://dfm.io/posts/histogram1/ Tue, 29 Jul 2014 00:00:00 +0000 https://dfm.io/posts/histogram1/ When I was writing my recent paper where I infer the population of exoplanets from an incomplete catalog of noisy measurements, I spent a lot of time thinking about histograms and I came across a few cute results that I wanted to share. Histograms are commonly used as simple “non-parametric” estimates of probability distributions but they get something of a bad rap (binning is sinning) because it’s easy to use them inappropriately! https://dfm.io/posts/python-c-extensions/ Fri, 03 Aug 2012 00:00:00 +0000 https://dfm.io/posts/python-c-extensions/ How to roll your own Python modules in C https://dfm.io/about/ Mon, 01 Jan 0001 00:00:00 +0000 https://dfm.io/about/ Hi, I’m Dan Foreman-Mackey I’m a Research Engineer at Anthropic in New York City, working on AI Interpretability. Previously, I worked on JAX at Google DeepMind, and before that I was an astrophysicist who developed open source scientific software. Code I write a lot of code for work and in my spare time. All my projects live in public repositories on GitHub. Here are some of my most popular research codes: