ftoolss 6.0.26

Small C extensions for local image filters (fmedian, fsigma)

ftoolss

High-performance C extensions for image processing and curve fitting.

Features

Export	Description
fmedian	Local median filter (2D/3D auto-dispatch)
fmedian2d, fmedian3d	Direct 2D/3D median
fsigma	Local ? filter (2D/3D auto-dispatch)
fsigma2d, fsigma3d	Direct 2D/3D sigma
fgaussian_f32, fgaussian_f64	Gaussian profile (Accelerate-optimized)
fmpfit_pywrap	Auto dtype dispatch (single spectrum)
fmpfit_f64_pywrap, fmpfit_f32_pywrap	Single spectrum fitting
fmpfit_block_pywrap	Auto dtype dispatch (block fitting)
fmpfit_f64_block_pywrap, fmpfit_f32_block_pywrap	Block fitting
MPFitResult	Named tuple for fit results

Common features: NaN handling, edge handling, optional center exclusion (fmedian/fsigma).

Installation

pip install .                        # Standard install
pip install -e .                     # Editable (development)
python setup.py build_ext --inplace  # Build extensions only

Requirements: Python 3.8+, NumPy ?1.20, C compiler. macOS Accelerate used when available.

Quick Start

import numpy as np
from ftoolss import fmedian, fsigma, fgaussian_f32, fgaussian_f64

# 2D/3D median and sigma filters
data = np.random.randn(100, 200)
median = fmedian(data, (3, 3), exclude_center=1)
sigma = fsigma(data, (3, 3), exclude_center=1)

# 3D works the same way
data_3d = np.random.randn(50, 100, 50)
median_3d = fmedian(data_3d, (3, 3, 3))

# Gaussian profiles
x = np.linspace(-10, 10, 1000, dtype=np.float32)
profile = fgaussian_f32(x, i0=1.0, mu=0.0, sigma=1.5)

Curve Fitting (fmpfit)

Single-spectrum fitting with auto dtype dispatch:

from ftoolss.fmpfit import fmpfit_pywrap
import numpy as np

x = np.linspace(-5, 5, 100)
y = 2.5 * np.exp(-0.5 * ((x - 1.0) / 0.8)**2) + np.random.randn(100) * 0.1

result = fmpfit_pywrap(
    deviate_type=0,  # Gaussian model
    parinfo=[
        {'value': 1.0, 'limits': [0.0, 10.0]},  # amplitude
        {'value': 0.0, 'limits': [-5.0, 5.0]},  # mean
        {'value': 1.0, 'limits': [0.1, 5.0]}    # sigma
    ],
    functkw={'x': x, 'y': y, 'error': np.ones_like(y) * 0.1}
)
print(f"Best-fit: {result.best_params}, ?²={result.bestnorm:.2f}")
print(f"SciPy-style errors (full Hessian): {result.xerror_scipy}")

Block fitting (multiple spectra in one call, GIL-free):

from ftoolss.fmpfit import fmpfit_block_pywrap
import numpy as np

n_spectra, mpoints, npar = 100, 200, 3
x = np.tile(np.linspace(-5, 5, mpoints), (n_spectra, 1))
y = ...  # shape (n_spectra, mpoints)
error = np.ones_like(y) * 0.1
p0 = np.tile([1.0, 0.0, 1.0], (n_spectra, 1))
bounds = np.array([[[0, 10], [-5, 5], [0.1, 5]]] * n_spectra)

results = fmpfit_block_pywrap(0, x, y, error, p0, bounds)
# results['best_params'] shape: (n_spectra, npar)

API Reference

fmedian / fsigma

fmedian(data, window_size, exclude_center=0)
fsigma(data, window_size, exclude_center=0)

• window_size: (x, y) or (x, y, z) ? odd positive integers
• exclude_center: 1 to exclude center from calculation (useful for outlier detection)
• Returns: float64 array, same shape as input

fgaussian_f32 / fgaussian_f64

Computes: i0 * exp(-((x - mu)² / (2 * sigma²))

fgaussian_f32(x, i0, mu, sigma)  # float32, fastest
fgaussian_f64(x, i0, mu, sigma)  # float64

fmpfit

Levenberg-Marquardt least-squares fitting. See src/ftoolss/fmpfit/README.md for details.

Function	Description
fmpfit_pywrap	Auto dtype dispatch (single spectrum)
fmpfit_f64_pywrap	float64 single spectrum
fmpfit_f32_pywrap	float32 single spectrum
fmpfit_block_pywrap	Auto dtype dispatch (block of spectra)
fmpfit_f64_block_pywrap	float64 block fitting
fmpfit_f32_block_pywrap	float32 block fitting

Examples

See examples/ directory for complete working examples.

Testing

pytest                              # Run all tests
pytest --cov=ftoolss --cov-report=html  # With coverage

Performance

Module	Speedup vs NumPy	Notes
fgaussian_f32	5-10×	Accelerate vvexpf
fgaussian_f64	2-3×	Accelerate vvexp
fmedian/fsigma	?	Sorting networks for small windows
fmpfit	?	GIL-free, 4× speedup with 6 threads

License

MIT

Credits

• Sorting networks: Bert Dobbelaere
• MPFIT: Craig Markwardt (cmpfit, MINPACK-1 derivative)