Largest-Triangle-Three-Buckets algorithm for downsampling time series-like data
Numpy implementation of Steinarsson’s Largest-Triangle-Three-Buckets algorithm for downsampling time series–like data while retaining the overall shape and variability in the data
LTTB is well suited to filtering time series data for visual representation, since it reduces the number of visually redundant data points, resulting in smaller file sizes and faster rendering of plots.
Note that it is not a technique for statistical aggregation, cf. regression models or non-parametric curve fitting / smoothing.
Install the lttb package into your (virtual) environment:
$ pip install lttb
The function lttb.downsample() can then be used in your Python code:
import numpy as np import lttb # Generate an example data set of 100 random points: # - column 0 represents time values (strictly increasing) # - column 1 represents the metric of interest: CPU usage, stock price, etc. data = np.array([range(100), np.random.random(100)]).T # Downsample it to 20 points: small_data = lttb.downsample(data, n_out=20) assert small_data.shape == (20, 2)
A test data set is provided in the source repo in tests/timeseries.csv. It was downloaded from http://flot.base.is/ and converted from JSON to CSV.
This is what it looks like, downsampled to 100 points:
By default, downsample() checks that the input data satisfies the following constraints:
it is a two-dimensional array of two columns;
the values in the first column are strictly increasing; and
there are no missing (NaN) values in the data.
These checks can be skipped (e.g. if you know that your data will always meet these conditions), or additional checks can be added (e.g. that the time values must be evenly spaced), by passing in a different list of validation functions, e.g.:
# No input validation: small_data = lttb.downsample(data, n_out=20, validators=) # Stricter check on x values: from lttb.validators import * small_data = lttb.downsample(data, n_out=20, validators=[has_two_columns, x_is_regular])
0.3.1 / 2020-10-14
All modules and functions now have docstrings.
[dev] The library is now also tested against Python 3.9. No changes to the code were required.
[dev] The project is now hosted on SourceHut; the links on the PyPI page have been updated.
[dev] CI testing has been migrated from Travis to builds.sr.ht.
0.3.0 / 2020-09-15
Validation of input data is now configurable.
New default: downsample() raises ValueError if input data contains NaN values. This can be disabled by removing contains_no_nans() from the list of validators.
[dev] Imports are now sorted with isort.
0.2.2 / 2020-01-08
setup.py was fixed so that this package can be installed in Python 2 again.
0.2.1 / 2019-11-25
[dev] Versions are now managed with setuptools_scm rather than bumpversion.
[dev] The code is formatted with Black.
0.2.0 / 2018-02-11
Released on PyPI (on 2019-11-06)
0.1.0 / 2017-03-18
JA Viljoen – original Numpy implementation
Guillaume Bethouart – performance improvements
Jens Krüger – fix for py27
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.