downsample: Collection of downsample algorithms for Python (Python using a C implementation)

This packages includes low level implementations written in C-Python of:

• The Largest Triangle Dynamic Buckets (LTD) downsampling algorithm
• The Largest Triangle Three Buckets (LTTB) downsampling algorithm
• The Largest Triangle One Bucket (LTOB) downsampling algorithm

The algorithm of LTTB was initially developed in (https://github.com/dgoeries/lttbc.git) Parts of this code have been translated and refers to the work of:

• Ján Jakub Naništa (https://github.com/janjakubnanista/downsample) (Typescript)
• Hao Chen (https://github.com/haoel/downsampling) (Go)

The algorithms are described in the work of Sveinn Steinarsson (https://github.com/sveinn-steinarsson/flot-downsample/).

Known features and requirements:

• The algorithm requires that x data is increasing and finite.
• y data must be finite; otherwise, issues may arise.
• x and y data must have the same length.
• The downsample algorithm returns a tuple of two arrays with data type double

Installing

You can also install it from PyPI to use in other environments with Python 3.10 or later:

pip install downsample

How to use on the field

All functions take an input for x and y in addition to the threshold:

from downsample import ltob, lttb, ltd
import numpy as np

array_size = 10000
threshold = 1000

x = np.arange(array_size, dtype=np.int32)
y = np.random.randint(1000, size=array_size, dtype=np.uint64)

x_l = x.tolist()
y_l = y.tolist()

for func in {ltd, ltob, lttb}:
    nx, ny = func(x, y, threshold)
    assert len(nx) == threshold
    assert len(ny) == threshold
    assert nx.dtype == np.double
    assert ny.dtype == np.double

    # List data or a mixture is accepted as well!
    nx, ny = func(x_l, y_l, threshold)
    assert len(nx) == threshold
    assert len(ny) == threshold
    assert nx.dtype == np.double
    assert ny.dtype == np.double

Performance Overview

For a performance overview, a sample of 7.500 data points was analyzed with a threshold set at 500 points. The performance test was conducted on a single core, utilizing a base clock speed of 3.70 GHz and 32 MB of L3 cache.

• LTD: 252.3 us
• LTOB: 51.6 us
• LTTB: 60.6 us