The SciPy 2016 conference was last week in Austin, Texas. Despite the suffocating heat, I got to visit the Lyndon B Johnson Presidential Library, see Ghostbusters in the Alamo Drafthouse, and sleep in an airstream trailer in someone’s back yard. More importantly, I had the opportunity to talk about pyglet_helper, and learn a lot of useful things about python. Here are a few personal highlights.
Making Python Packages for Windows is Hard
Frequent visitors to PyPI may have noticed that many common packages have recently switched to using wheels. No longer will we twiddle our thumbs restlessly as we watch setup.py install churn through line after line of gcc warnings! That is due to the excellent efforts of Nathaniel Smith, who spoke about his work:
Wheels can be built for several different versions of python using a provided docker image. The Numpy and Scipy builds are built with OpenBLAS, which is an exciting development. Smith seems confident that it will be possible to bundle OpenBLAS with Numpy wheels on Windows soon as well, which will be a big time-saver for me.
At 9:12, he describes exactly what is needed to easily compile wheels for Windows - the lack of a good compiler. This was brought up by an audience member at the python core developers panel at Euro Python a few days ago and the panellists seemed surprised. I think the problem is that the people with the domain knowledge of compilers are exclusively linux users and windows users would prefer to use Anaconda. In between are a few like me, who both need to use Windows for device drivers but also want to use obscure packages in PyPI, without fussing with the visual studio compiler.
As much joy as writing this blog has been, it has forced me to confront the fact that I am not a good writer and I should be better at editing my text. I acquired an appreciation for style guides from the one journalism class I took in undergrad, but 10 years later I’ve forgotten most of the rules. Thus, I am keenly interested in Michael Pacer and Jordan Suchow’s project, proselint:
I’ve tried it out for myself on my blog posts, and at the moment it has difficulty with markdown, especially in accidentally prose-linting code blocks. Regular expressions have only a limited ability to fix bad writing, but I will take all the help I can get. I’m interested in building a PyCharm extension for proselint and in incorporating proselint into my githubio build process. I’m not sure how successful I will be in my current set-up, as my build uses ruby.
Though I have been working in experimental physics for the past six years, I had never heard of the Experimental Physics and Industrial Control System (EPICS) until Daniel Allen’s talk. I work on experiments with only up to one control computer at a time, whereas EPICS is used big, hundred-party experiments such as Synchrotrons or LIGO. From a conceptual level, it reminds me of the Robotics Operating System, with its publishers and subscribers. In his talk, Allan describes an experimentalists’ dream: capture your data along with the operations in a retrievable format, and have a system that can roll back to a recoverable state if something goes wrong.
These are design goals for my projects as well, even if they only have 3 users. I have limited time with my equipment as well, and needing to repeat an experiment because some metadata was missing is a waste. I don’t quite understand the usage of asyncio in this context, but I’m eager to learn how it could benefit my experimental automation.
Imputation is Important for the Real World
John’s recently defended PhD thesis is about applying machine learning techniques to filling missing data (imputation) on incomplete ranked ballots in order to maximize the fairness of elections. Imputation is an important consideration when applying algorithms to data in the real world, as Deborah Hanus demonstrated in her talk, featuring her work with optimizing the treatments for HIV patients:
Due to issues surrounding shift work and childcare, transportation access, or human error, HIV patients sometimes miss their regularly scheduled blood tests. Imputation allows the machine learning algorithm to predict optimal dosages even in the presence of missing data. Aside from being interesting research, the talk also provides a good layman’s explanation of imputing random walks over time.