Python function for performing a linear binning that is optimized for sparsity
Project description
sparse_linear_binning: linear binning (optimized for sparsity)
Performs a linear binning technique described in Wand and Jones on a regularlyspaced grid in an arbitrary number of dimensions. The asymptotic behavior of this binning technique performs better than socalled simple binning (i.e. as in histograms). Each data point in ddimensional space must have an associated weight (for equally weighted points just use a weight of 1.0 for each point).
For example, within a (segment of a) 2D grid with corners A, B, C, and D and a 2D data point P with weight wP:
AB             P                 DC
 Assign a weight to corner A of the proportion of area between P and C (times wP)
 Assign a weight to corner B of the proportion of area between P and D (times wP)
 Assign a weight to corner C of the proportion of area between P and A (times wP)
 Assign a weight to corner D of the proportion of area between P and B (times wP)
Note that the under and overflow bins need to be accounted for when specifying the numbers of grid points in each dimension (grid points act as bin centers). For instance, if you want grid points in steps of 0.1 in a range of [0,1] (i.e. (0, .1, .2, .3, .4, .5, .6, .7, .8, .9, 1)), specify the number of grid points to be 11. Internally, the grid points are stored in a high performance, C++based hash table (sparsepp). This allows for finer binning in some circumstances because the hash table doesn’t allocates memory for grid points with nearzero weight. To accommodate arbitrary numbers of bins along each dimension, an arbitrary precision numeric library (boost multiprecision) may be used internally and will negatively impact performance. If this degradation in performance is unacceptable, consider reducing the number of grid points in such a way that the product of grid points in all dimensions is less than the numeric maximum of “unsigned long long” on your system. For instance, in 20 dimenisons with each dimension having 51 grid points gives a total of 14171098670753043575626125424226001 potential grid points at which point the arbitrary precision library must take care of all arithmetic related to determining grid points.
Quickstart
 pip install sparse_linear_binning
or
 git clone https://github.com/jhetherly/sparse_linear_binning
 cd sparse_linear_binning
 python setup.py install
Example
This constructs one million random 2D points in the unit square with random weights and constructs a grid of 51 by 51 (can be different along different dimensions) linearly binned “bin centers.” The boundaries of the grid of bin centers are specified by extents and can be thought of as the under and overflow bins (i.e. these are the coordinates of the first and last bin centers).
from sparse_linear_binning import sparse_linear_binning import numpy as np # generate one million random 2D points and weights # (should take less than a second to bin) n_samples=1000000 D=2 # coordinates, weights, and extents must be of type "double" sample_coords = np.random.random(size=(n_samples, D)) sample_weights = np.random.random(size=n_samples) extents = np.tile([0., 1.], D).reshape((D, 2)) n_bins = np.full(D, 51) coords, weights = sparse_linear_binning(sample_coords, sample_weights, extents, n_bins) # check that weights on grid match original weights print(np.allclose(weights.sum(), sample_weights.sum()))
Dependencies
 numpy
# 1.1.0
 Added support for automatically converting Numpy arrays to proper dtypes.
 Added more explicit tests.
Project details
Release history Release notifications
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Filename, size  File type  Python version  Upload date  Hashes 

Filename, size sparse_linear_binning1.1.0.tar.gz (1.3 MB)  File type Source  Python version None  Upload date  Hashes View hashes 
Hashes for sparse_linear_binning1.1.0.tar.gz
Algorithm  Hash digest  

SHA256  bdf33e1b799b66c53b057647f3ddd09dcd898be79b52f516247843f499615f46 

MD5  73bfd19973feb5d7705f9ac2c9768e55 

BLAKE2256  2788c9e113c4428b0c105c3e94f947180d1f2a09a0a7a662902a0eebba8a905d 