zfit: Scalable Pythonic Model Fitting for High Energy Physics

0.28.0 · active · verified Sun Apr 12

zfit is a modern, scalable Python library for statistical model fitting, primarily designed for High Energy Physics but applicable broadly. It leverages TensorFlow (and optionally JAX) for accelerated computations on CPUs and GPUs, offering a user-friendly API for defining, manipulating, and fitting complex probability density functions. The current version is 0.28.0, with a release cadence of typically a few months between minor versions, often driven by new feature development and backend compatibility updates.

Warnings

breaking The `effsize` weight correction parameter was renamed to `sumw2` in certain contexts for asymptotic uncertainty calculations.
Fix: Update calls using `effsize` to `sumw2`.
gotcha The default padding for KDE (Kernel Density Estimation) functions, specifically `KDE1DimExact`, changed from `False` to `0.1`.
Fix: If exact reproduction of pre-0.27.0 KDE behavior at boundaries is critical, explicitly set `padding=False` when creating KDEs.
gotcha zfit frequently updates its TensorFlow and TensorFlow-Probability backend requirements. Major version bumps of these dependencies can cause environment conflicts or necessitate upgrades.
Fix: Always check the `requires_python` and the specific backend requirements in the release notes or `pyproject.toml`. Use `pip install zfit[tf]` or `zfit[jax]` to ensure compatible backend versions are installed, or manage your environment carefully if pinning specific backend versions.
breaking As of v0.27.0, zfit explicitly requires Python 3.10 or newer. Support for Python 3.13 was added in v0.28.0.
Fix: Ensure your Python environment is at least 3.10. For the latest features and compatibility, use Python 3.10-3.13.

Install

pip install zfit[tf] Recommended installation with TensorFlow backend
pip install zfit[jax] Installation with JAX backend
pip install zfit Minimal installation (includes CPU TensorFlow by default)

Imports

Parameter
```
from zfit import Parameter
```
Space
```
from zfit import Space
```
pdf
```
import zfit.pdf as zpdf
```
Commonly aliased for brevity, e.g., zfit.pdf.Gauss
data
```
import zfit.data as zdata
```
Commonly aliased for brevity, e.g., zfit.data.Data
loss
```
import zfit.loss as zloss
```
Commonly aliased for brevity, e.g., zfit.loss.UnbinnedNLL
minimize
```
import zfit.minimize as zmin
```
Commonly aliased for brevity, e.g., zfit.minimize.Minuit

Quickstart

This quickstart demonstrates defining parameters, an observable space, creating a Gaussian Probability Density Function (PDF), generating data from it, setting up an unbinned negative log-likelihood loss, and performing a fit using the Minuit minimizer. It's a fundamental workflow for zfit.

import zfit
from zfit import z

# Define parameters
mu = zfit.Parameter('mu', 0.0, -1.0, 1.0)
sigma = zfit.Parameter('sigma', 1.0, 0.1, 10.0)

# Define the observable space
obs = zfit.Space('x', limits=(-5, 5))

# Create a Gaussian PDF
gauss = zfit.pdf.Gauss(mu=mu, sigma=sigma, obs=obs)

# Generate some data
data = gauss.sample(n=1000)

# Create a likelihood loss function
nll = zfit.loss.UnbinnedNLL(model=gauss, data=data)

# Create a minimizer
minimizer = zfit.minimize.Minuit()

# Perform the fit
result = minimizer.minimize(nll)

# Print the fit results
print(result.params)

view raw JSON →