On Apr 21, 2018 02:41, "Matthew Brett" <matthew.brett@gmail.com> wrote: Hi, On Sat, Apr 21, 2018 at 1:12 AM, Warren Weckesser <warren.weckesser@gmail.com> wrote:
On Fri, Apr 20, 2018 at 9:22 AM, Eric Larson <larson.eric.d@gmail.com> wrote:
Do any of you have suggestions for issues we could work on? I'd love to hear any suggestions.
One option would be to suggest that each interested participant open the SciPy issues and PRs list, and then sub-select based on the `label` corresponding to a submodule they might be interested in. That way people will gravitate toward domains where they have knowledge and feel comfortable, and it cuts down on the number of issues they need to look
at
(because we have so many).
And of course there is also the "good first issue" label, which can be useful even if it's inconsistently applied.
One of those issues is https://github.com/scipy/scipy/issues/7168. There are hundreds of function in the library that do not have an "Examples" section. It would be great to bring that number down. Adding some examples could be a good way to get folks started with the SciPy development work flow (building the library, building the docs, and making pull requests). If you pursue this, be sure that everyone is aware of PEP 8 (https://www.python.org/dev/peps/pep-0008/) and knows where to find the docstring standard (https://numpydoc.readthedocs.io/en/latest/format.html#docstring-standard ).
Excellent suggestion, thanks. I'll offer that as a fall-back, and maybe as a first hurdle, before trying something with more meat. Cheers, Matthew A pet peeve of mine, but I know there has been some work on getting the scipy.signal.fftconvolve function to broadcast across dimensions [1], instead of only working with vectors. However, it has never been completed. This is a particularly important case because it is a necessary prerequisite for an implementation of the overlap-add convolution algorithm [2], one of the more critical basic signal processing algorithms that scipy still lacks. It \provides a massive performance boost when you have two large (both >= ~500 samples) but different-sized vectors. So although getting the overlap-add method implemented is probably too much, just allowing applying the 1D fftconvolve across dimensions might be within in reason. The advantage for the hackathon is that it provides the groundwork for a big improvement, but shouldn't require any serious signal-processing knowledge, since the actual algorithm is already implemented. This just requires implement the dimension-handling logic. [1] https://github.com/scipy/scipy/issues/3525 [2] https://en.wikipedia.org/wiki/Overlap-add_method