I see usefulness in both Gitter and GH Discussions:
- Gitter for issues that require some back-and-fourth for a short amount of time and are likely very user-specific (e.g.: my compilation fails and I don't know how to fix it), if that gets lost in the history it is probably ok since other users may not have the same question
- Discussions for questions that lots of users may have (and then in the future find there answered, like StackOverflow), or questions that take a long time to answer (to have one consistent thread there instead of being spread out in Gitter between other topics). I also think the "Show and tell" category could be quite nice, and a message like that would disappear quickly in Gitter. And I personally like the idea of everything in one place (repository) a lot.
I’d see gitter as more for announcements and for developer discussions (scroll back in the history, it’s pretty much the same 4-5 people with maybe a couple of user questions), while most users usually open an issue (there are more question issues than questions in gitter, I believe). Discussions gives them a safer, nicer place to open these things that hopefully won’t be as large of a barier, and will be more discoverable after they are answered. But they should be seen mostly as better orginization and a few extra features like threads and answers on top of issues.
I haven't had the chance to use discussions much in mplhep yet, pretty quiet these days, but I am a big supporter. It's a great place for "not quite issue"'s. Gitter still works when one has a question that's more random, but it might as well not keep history past few days
My histogram is having a Regular and a strCategory axis and I want to show the cumulative sum of the categories, but with the different components stacked on top of each other.
I found out I can just use np.cumsum(axis=<category axis>) to calculate the cumulative sum and create a new histogram from that. In had also managed to draw a stacked histogram by plotting multiple plt.bar with different bottom=... arguments taken from the cumulative sums. But I dislike that visualy, because I want to have a black outline for my components, but no outlines for the individual bins, as seen in the screenshot.
though admittedly i have not tested is super thoroughly with bh, but as especially if exported to numpy arrays it should be trivial to plot stuff like this
(And the correct way to do it will be the function in mplhep, or possibly mpl’s new stairs plot)
Will probably bump minor version for mplhep with it
Hi, I used to be able to modify histogram contents like this
import boost_histogram as bh
import numpy as np
bins = [0, 1, 2]
hist = bh.Histogram(bh.axis.Variable(bins), storage=bh.storage.Weight())
yields = [3, 4]
var = [0.1, 0.2]
hist[...] = np.stack([yields, var], axis=-1)
hist.view().value /= 2That worked until boost-histogram version 0.11.1. In the current master it does not anymore, I think scikit-hep/boost-histogram#475 changes the behavior:
Traceback (most recent call last):
File "test.py", line 10, in <module>
hist.view().value /= 2
File "[...]/boost_histogram/_internal/view.py", line 57, in fset
self[name] = value
File "[...]/boost_histogram/_internal/view.py", line 49, in __setitem__
raise ValueError("Needs matching ndarray or n+1 dim array")
ValueError: Needs matching ndarray or n+1 dim arrayIs there another way to achieve this now?
hist.view() /= 2.
Boost-histogram 0.12 is out. It’s mostly a bugfix release, fixing several important bugs, with a few minor things. The current develop branch has PlottableProtocol, which needs a little more time before being released, but feel free to try it out if you like the bleeding edge!
Hi all, there is currently a push in scipy to include Boost as a dependency. This is initiated by the wish to replace current implementations in scipy.special and scipy.stats with those in Boost.Math. Someone else (not me!) then brought up that one could then also base histograms and the scipy.stats.binned_statistics on Boost.Histogram.
I then told them about boost-histogram and the possible performance increases. Not anything I could work on anytime soon, but exciting to see interest in Scipy about this sort of thing.
Someone also mentioned the problem with incremental filling, which is currently not efficiently possible with numpy.histogram, which boost-histogram solves.
I was told that scipy.stats.binned_statistics is a pure python implementation currently, so basing it on Boost.Histogram would surely speed up that code.
Sorry, currently, it is impossible. Boost.Histogram requires C++14, and SciPy ships manylinux1 wheels. You can’t build the C++14 code in Boost.Histogram with GCC 4.8.2. However, the PyPA plans to drop manylinux1 this summer, so at some point this year, in theory, SciPy/NumPy/etc. will all stop shipping manylinux1 wheels, and suddely the minimum compiler anyone cares about will be much higher (8.3.1 for manylinux2010), making this possible for the first time!
It’s going to be a big deal, because pip 9 can’t download manylinux2010 wheels, which is the main reason they still ship manylinux1 wheels today. And recently I found if you use Ubuntu 18.04, and use the
python-3.8 package, you still get pip 9 even on Python 3.8!!! (which is totally unsupported). I hate distro packaging sometimes…
They are currently debating whether they want to depend on Boost at all (although this seems to have some support...)
Boost.Histogram was only brought up because Boost.Math is so intertwined with other Boost libs that it is not feasible to extract only Boost.Math, they then have to depend on all of Boost anyway.
Interestingly, the point about C++14 may also come up when they start to use Boost.Math. The implementations in Boost.Math depend on varying versions of the C++ standard, because the maintainers leave it to the contributors which standard they prefer (which is not the best idea IMHO)
