pygram11
Simple and fast histogramming in Python accelerated with OpenMP with help from pybind11.
pygram11 provides functions for very fast histogram calculations (and the variance in each bin) in one and two dimensions. The API is very simple; documentation can be found here (you'll also find some benchmarks there).
Installing
From PyPI
Binary wheels are provided for Linux and macOS. They can be installed from PyPI via pip:
pip install pygram11
From conda-forge
For installation via the conda package manager pygram11 is part of conda-forge.
conda install pygram11 -c conda-forge
Please note that on macOS the OpenMP libraries from LLVM (libomp) and Intel (libiomp) may clash if your conda environment includes the Intel Math Kernel Library (MKL) package distributed by Anaconda. You may need to install the nomkl package to prevent the clash (Intel MKL accelerates many linear algebra operations, but does not impact pygram11):
conda install nomkl ## sometimes necessary fix (macOS only)
From Source
You need is a C++14 compiler and OpenMP. If you are using a relatively modern GCC release on Linux then you probably don't have to worry about the OpenMP dependency. If you are on macOS, you can install libomp from Homebrew (pygram11 does compile on Apple Silicon devices with Python version >= 3.9 and libomp installed from Homebrew). With those dependencies met, simply run:
git clone https://github.com/douglasdavis/pygram11.git --recurse-submodules
cd pygram11
pip install .
Or let pip handle the cloning procedure:
pip install git+https://github.com/douglasdavis/[email protected]
Tests are run on Python versions >= 3.7 and binary wheels are provided for those versions.
In Action
A histogram (with fixed bin width) of weighted data in one dimension:
>>> rng = np.random.default_rng(123)
>>> x = rng.standard_normal(10000)
>>> w = rng.uniform(0.8, 1.2, x.shape[0])
>>> h, err = pygram11.histogram(x, bins=40, range=(-4, 4), weights=w)
A histogram with fixed bin width which saves the under and overflow in the first and last bins:
>>> x = rng.standard_normal(1000000)
>>> h, __ = pygram11.histogram(x, bins=20, range=(-3, 3), flow=True)
where we've used __ to catch the None returned when weights are absent. A histogram in two dimensions with variable width bins:
>>> x = rng.standard_normal(1000)
>>> y = rng.standard_normal(1000)
>>> xbins = [-2.0, -1.0, -0.5, 1.5, 2.0, 3.1]
>>> ybins = [-3.0, -1.5, -0.1, 0.8, 2.0, 2.8]
>>> h, err = pygram11.histogram2d(x, y, bins=[xbins, ybins])
Manually controlling OpenMP acceleration with context managers:
>>> with pygram11.omp_disabled(): # disable all thresholds.
... result, _ = pygram11.histogram(x, bins=10, range=(-3, 3))
...
>>> with pygram11.omp_forced(key="thresholds.var1d"): # force a single threshold.
... result, _ = pygram11.histogram(x, bins=[-3, -2, 0, 2, 3])
...
Histogramming multiple weight variations for the same data, then putting the result in a DataFrame (the input pandas DataFrame will be interpreted as a NumPy array):
>>> N = 10000
>>> weights = pd.DataFrame({"weight_a": np.abs(rng.standard_normal(N)),
... "weight_b": rng.uniform(0.5, 0.8, N),
... "weight_c": rng.uniform(0.0, 1.0, N)})
>>> data = rng.standard_normal(N)
>>> count, err = pygram11.histogram(data, bins=20, range=(-3, 3), weights=weights, flow=True)
>>> count_df = pd.DataFrame(count, columns=weights.columns)
>>> err_df = pd.DataFrame(err, columns=weights.columns)
I also wrote a blog post with some simple examples.
Other Libraries
- boost-histogram provides Pythonic object oriented histograms.
- Simple and fast histogramming in Python using the NumPy C API: fast-histogram (no variance or overflow support).
- To calculate histograms in Python on a GPU, see cupy.histogram.
If there is something you'd like to see in pygram11, please open an issue or pull request.
150 Nov 03, 2022
209 Sep 19, 2022
1 Jul 12, 2022
2 Jun 15, 2022
3.8k Jan 02, 2023
145 Dec 20, 2022
4 Oct 10, 2022
9.4k Jan 07, 2023
9 Dec 13, 2022
2 Oct 30, 2022
31 Dec 15, 2022
29 Jul 05, 2022
3.3k Dec 31, 2022
719 Jan 04, 2023
192 Dec 16, 2022
1.7k Jan 07, 2023
6.7k Jan 09, 2023
258 Nov 22, 2022
1k Jan 01, 2023