Changepoint detection with Pruned Exact Linear Time.

Usage

Python

from pelt import predict

predict(signal, penalty=20, segment_cost_function="l1", jump=10, minimum_segment_length=2)

Rust

use std::num::NonZero;
use pelt::{Pelt, SegmentCostFunction};

// Setup the structure for calculating changepoints
let pelt = Pelt::new()
  .with_jump(NonZero::new(5).expect("Invalid number"))
  .with_minimum_segment_length(NonZero::new(2).expect("Invalid number"))
  .with_segment_cost_function(SegmentCostFunction::L1);

// Do the calculation on a data set
let penalty = 10.0;
let result = pelt.predict(&signal[..], penalty)?;

Run locally

# Install maturin inside a Python environment
python3 -m venv .env
source .env/bin/activate
pip install maturin numpy

# Create a Python package from the Rust code
maturin develop

# Open an interpreter
python

>>> from pelt import predict
>>> import numpy as np
>>> signal = np.array([np.sin(np.arange(0, 1000, 10))]).transpose()
>>> predict(signal, penalty=20)

Benchmarks

Warning

Like all benchmarks, take these with a grain of salt.

Python

Comparison with ruptures:

Cost Function	Data Points	Data Dimension	Mean pelt	Mean ruptures	Times Faster
L2	100	1D	2.042 μs	3.110 ms	1523.1x
L2	100	2D	2.420 μs	3.148 ms	1300.9x
L2	1000	1D	106.783 μs	190.108 ms	1780.3x
L2	1000	2D	55.815 μs	104.357 ms	1869.7x
L2	10000	1D	20.407 ms	12.859 s	630.2x
L2	10000	2D	2.317 ms	1.797 s	775.6x
L1	100	1D	11.245 μs	5.041 ms	448.3x
L1	100	2D	21.587 μs	5.350 ms	247.8x
L1	1000	1D	321.619 μs	187.656 ms	583.5x
L1	1000	2D	2.186 ms	628.126 ms	287.3x
L1	10000	1D	13.215 ms	15.615 s	1181.6x
L1	10000	2D	84.453 ms	30.508 s	361.2x

Command

maturin develop
python benches/bench_compare.py

Profile

Command

cargo build --example simple --profile profiling \
 && samply record target/profiling/examples/simple tests/signals-large.csv

Credits

• fastpelt
• ruptures