gfit 0.4

Fast multi-gaussian curve fitting

gfit

Fast multi-Gaussian curve fitting for spectra and other 1D signals.

gfit fits many independent signals at once by decomposing each into a sum of symmetric or asymmetric Gaussian peaks. It is aimed at hyperspectral and spectroscopy workflows, but works for any regularly sampled 1D data.

Features

• Fit symmetric (amplitude, centre, width) or asymmetric (separate left/right widths) Gaussians
• Batch fitting over (n_spectra, n_points) arrays
• Automatic peak-based initialisation via initialise()
• Optional Numba JIT acceleration (pip install "gfit[all]")
• Pure-Python fallback when Numba is unavailable
• Multiprocessing for large batches; automatic single-threaded fallback in Pyodide / WebAssembly

Installation

Minimal install (NumPy and tqdm only; pure-Python fallback):

pip install gfit

Recommended install (NumPy, SciPy, tqdm, Numba):

pip install "gfit[all]"

SciPy only (no Numba; slower fitting, no JIT):

pip install "gfit[scipy]"

On zsh, quote extras: pip install "gfit[all]" (unquoted brackets are treated as globs).

Gaussian fitting (gfit, initialise, evaluate) imports without SciPy, but calling the fitters raises an ImportError with install instructions unless SciPy is installed.

Quick start

import numpy as np
from gfit import gfit, initialise, evaluate
from gfit.util import split_coeff, rand_signal

# x-axis and a batch of noisy spectra (n_spectra, n_points)
x = np.linspace(-10, 10, 200)
X = np.array([rand_signal(x, snr=14)[0] for _ in range(100)])

# Fit 3 symmetric Gaussians per spectrum
n = 3
x0 = initialise(x, X, n, sym=True, d=4)
params = gfit(x, X, x0, n, sym=True, nthreads=-1, vb=True)

# Split fitted parameters into amplitude, centre, width arrays
a, b, c = split_coeff(params, sym=True)

# Evaluate the fitted model for one spectrum
y_model = evaluate(x, params[0:1], sym=True)[0]

API overview

Function	Description
gfit(x, X, ...)	Fit Gaussians to a batch of spectra
initialise(x, X, n, ...)	Estimate starting parameters from local peaks
evaluate(x, M, sym=...)	Evaluate one or more multi-Gaussian models

gfit parameters

Parameter	Default	Meaning
x	—	1D array of sample positions
X	—	2D array (n_spectra, n_points) of values to fit
x0	None	Initial parameters; computed via initialise() if omitted
n	3	Number of Gaussian components per spectrum
sym	True	Symmetric (True) or asymmetric (False) peaks
thresh	-1	Skip fitting if peak amplitude below threshold; -1 fits all
nthreads	-1	Parallel workers; 1 forces single-threaded
vb	True	Show tqdm progress bar

Additional keyword arguments (ftol, xtol, maxiter, …) are passed to scipy.optimize.least_squares.

Parameter layout

Fitted coefficients are stored as a flat array per spectrum:

• Symmetric: [a₀, b₀, c₀, a₁, b₁, c₁, …] — amplitude, centre, width
• Asymmetric: [a₀, b₀, c₁₀, c₂₀, …] — amplitude, centre, left width, right width

Use gfit.util.split_coeff() and stack_coeff() to convert between flat arrays and separate parameter lists.

Utilities

from gfit.util import remove_hull, get_bounds, benchmark

# Remove a convex hull baseline before fitting
X_corrected = remove_hull(X, upper=True, div=True)

# Bounds derived from an initial guess (also used internally by gfit)
bounds = get_bounds(x, x0, sym=True)

# Simple performance check
benchmark(size=100, res=100, nthreads=1)

Lower-level single-spectrum fitters are available in gfit.internal:

from gfit.internal import fit_mgauss, fit_amgauss

Performance

• With Numba (gfit[all]): hot loops are JIT-compiled; batch initialisation can use parallel prange.
• With a minimal install or missing Numba: the same code runs as pure Python (correct, but slower). A warning is emitted once at import.
• nthreads=-1 (default) uses multiprocessing to fit spectra in parallel on desktop Python.
• nthreads=1 uses a single process (recommended for small batches or constrained environments).

License

MIT — see LICENSE.

Author

Sam Thiele — github.com/samthiele/gfit