The fifth scikits.learn coding sprint took place Friday, April 1st 2011. For anyone who is not familiar with it, scikits.learn is a fast and easy to use machine learning toolkit for the pylab environment (Python, NumPy, SciPy, Matplotlib.)
This was a good opportunity for me to get code reviews by the developers in order to bring my NMF code up to standards, so that it can be merged. Though I live far from every nucleus of scikits-learn developers, I efficiently participated via IRC. This way, I also got the chance to help out a bit on Mathieu Blondel’s Kernel PCA code, which will also be merged into main soon.
How it felt like¶
Short answer: awesome!
Slightly longer answer: Everybody was very encouraging and helpful. They gave me a lot of feedback from which I learned a lot, and they manifested the intention to merge soon. It is a pleasure to work on projects that you like and use, especially when the projects leaders and collaborators are so good to work with.
But the main reason why it makes me feel so good is that I’m proud to able to contribute on a project that I consider very significant and the best in the field from many points of view.
What I got done¶
Most of my work was on the non-negative matrix factorization module that I began some while back, but only intermitently worked on. It is now a solid module with high test coverage, documentation, and a cool simple example showing a sparse set of features for the digits dataset in scikits.learn. Apart from all the minor fixes in overall code quality and cleanliness, probably what is the most relevant is the improvement and the study of the initialization methods. I will look into this further and document it on this blog, the point is that the choice of initialization method greatly influences the speed of convergence, and in the case of a high-tolerance setting, also the error obtained. Some initializations are more fit for sparsity settings, while others are more fit for dense settings.
I have a theory that I plan to test out, regarding the use of different initialization methods for components and for data in a sparse setting.
What I learned¶
I think my greatest improvement was in terms of workflow and efficiency. While my code was under review, I was receiving frequent comments on my git pull request, and eventually I ended up responding to some comments even before they were posted :). I sent small fixes as pull requests to help other developers as much as I could. Before scikits.learn I had never worked on a project with so many developers, and I think I handled it well, even though I asked once or twice on the IRC channel for pieces of git-fu.
I learned that it’s difficult to tweak matplotlib subplots! I’m still staring at Alexandre Gramfort’s tweak in my example and I have no idea what he did to make it look so good. But I’ll figure it out soon, I’m sure.
I also learned a lot more about the intricacies of the scikits.learn APIs, the philosophy of ease of use, and the project tree in general.
In short, the coding sprint has been a great and rewarding experience, for which I thank all of you guys there!