zfit 0.5.5

scalable pythonic model fitting for high energy physics

zfit: scalable pythonic fitting

zfit is a highly scalable and customizable model manipulation and fitting library. It uses TensorFlow as its computational backend and is optimised for simple and direct manipulation of probability density functions. The project is affiliated with and well integrated into Scikit-HEP, the HEP Python ecosystem.

• Tutorials: Interactive IPython Tutorials

• Quick start: Example scripts

• Documentation: See stable documentation or latest documentation

• Questions: see the FAQ, ask on StackOverflow (with the zfit tag) or contact us directly.

• Physics/HEP: zfit-physics is the place to contribute/find more HEP related content

If you use zfit in research, please consider citing.

N.B.: zfit is currently in beta stage, so while most core parts are established, some may still be missing and bugs may be encountered. It is, however, mostly ready for production, and is being used in analyses projects. If you want to use it for your project and you are not sure if all the needed functionality is there, feel free to contact.

Why?

The basic idea behind zfit is to offer a Python oriented alternative to the very successful RooFit library from the ROOT data analysis package that can integrate with the other packages that are part if the scientific Python ecosystem. Contrary to the monolithic approach of ROOT/RooFit, the aim of zfit is to be light and flexible enough to integrate with any state-of-art tools and to allow scalability going to larger datasets.

These core ideas are supported by two basic pillars:

• The skeleton and extension of the code is minimalist, simple and finite: the zfit library is exclusively designed for the purpose of model fitting and sampling with no attempt to extend its functionalities to features such as statistical methods or plotting.

• zfit is designed for optimal parallelisation and scalability by making use of TensorFlow as its backend. The use of TensorFlow provides crucial features in the context of model fitting like taking care of the parallelisation and analytic derivatives.

How to use

While the zfit library provides a model fitting and sampling framework for a broad list of applications, we will illustrate its main features with a simple example by fitting a Gaussian distribution with an unbinned likelihood fit and a parameter uncertainty estimation.

Example in short

obs = zfit.Space('x', limits=(-10, 10))

# create the model
mu    = zfit.Parameter("mu"   , 2.4, -1, 5)
sigma = zfit.Parameter("sigma", 1.3,  0, 5)
gauss = zfit.pdf.Gauss(obs=obs, mu=mu, sigma=sigma)

# load the data
data_np = np.random.normal(size=10000)
data = zfit.Data.from_numpy(obs=obs, array=data_np)

# build the loss
nll = zfit.loss.UnbinnedNLL(model=gauss, data=data)

# minimize
minimizer = zfit.minimize.Minuit()
result = minimizer.minimize(nll)

# calculate errors
param_errors = result.error()

This follows the zfit workflow

Full explanation

The default space (e.g. normalization range) of a PDF is defined by an observable space, which is created using the zfit.Space class:

obs = zfit.Space('x', limits=(-10, 10))

To create a simple Gaussian PDF, we define its parameters and their limits using the zfit.Parameter class.

# syntax: zfit.Parameter("any_name", value, lower, upper)
  mu    = zfit.Parameter("mu"   , 2.4, -1, 5)
  sigma = zfit.Parameter("sigma", 1.3,  0, 5)
  gauss = zfit.pdf.Gauss(obs=obs, mu=mu, sigma=sigma)

For simplicity, we create the dataset to be fitted starting from a numpy array, but zfit allows for the use of other sources such as ROOT files:

mu_true = 0
sigma_true = 1
data_np = np.random.normal(mu_true, sigma_true, size=10000)
data = zfit.Data.from_numpy(obs=obs, array=data_np)

Fits are performed in three steps:

1. Creation of a loss function, in our case a negative log-likelihood.

2. Instantiation of our minimiser of choice, in the example the Minuit.

3. Minimisation of the loss function.

# Stage 1: create an unbinned likelihood with the given PDF and dataset
nll = zfit.loss.UnbinnedNLL(model=gauss, data=data)

# Stage 2: instantiate a minimiser (in this case a basic minuit)
minimizer = zfit.minimize.Minuit()

# Stage 3: minimise the given negative log-likelihood
result = minimizer.minimize(nll)

Errors are calculated with a further function call to avoid running potentially expensive operations if not needed:

param_errors = result.error()

Once we’ve performed the fit and obtained the corresponding uncertainties, we can examine the fit results:

print("Function minimum:", result.fmin)
print("Converged:", result.converged)
print("Full minimizer information:", result.info)

# Information on all the parameters in the fit
params = result.params
print(params)

# Printing information on specific parameters, e.g. mu
print("mu={}".format(params[mu]['value']))

And that’s it! For more details and information of what you can do with zfit, checkout the latest documentation.

Prerequisites

zfit works with Python versions 3.6 and 3.7. The following packages (amongst others) are required:

• tensorflow >= 2.x

• tensorflow_probability

• scipy >=1.2

• uproot

• iminuit

… and some minor packages. For a full list, check the requirements.

Installing

zfit is available on conda-forge and pip. If possible, use a conda or virtual environment and do:

For conda:

$ conda install zfit -c conda-forge

For pip (if you don’t use conda):

$ pip install zfit

For the newest development version, you can install the version from git with

$ pip install git+https://github.com/zfit/zfit

Contributing

Any idea of how to improve the library? Or interested to write some code? Contributions are always welcome, please have a look at the Contributing guide.

Contact

You can contact us directly:

• via e-mail: [e-mail]

• join our Gitter channel

Original Authors

Jonas Eschle <[e-mail]>

Albert Puig <[e-mail]>

Rafael Silva Coutinho <[e-mail]>

See here for all authors and contributors

Acknowledgements

zfit has been developed with support from the University of Zurich and the Swiss National Science Foundation (SNSF) under contracts 168169 and 174182.

The idea of zfit is inspired by the TensorFlowAnalysis framework developed by Anton Poluektov and TensorProb by Chris Burr and Igor Babuschkin using the TensorFlow open source library and more libraries.

Changelog

Develop

Major Features and Improvements

Breaking changes

Depreceations

Bug fixes and small changes

Experimental

Requirement changes

Thanks

0.5.5 (20.10.2020)

Upgrade to TensorFlow 2.3 and support for weighted hessian error estimation.

Added a one dimensional Convolution PDF

Major Features and Improvements

• upgrad to TensorFlow 2.3

• new Poisson PDF

Breaking changes

Depreceations

Bug fixes and small changes

• print parameter inside function context works now correctly

Experimental

• Computation of the covariance matrix and hessian errors with weighted data

• Convolution PDF (FFT in 1Dim) added (experimental, feedback welcome!)

Requirement changes

• TensorFlow==2.3 (before 2.2)

• tensorflow_probability==0.11

• tensorflow-addons # spline interpolation in convolution

Thanks

0.5.4 (16.07.2020)

Major Features and Improvements

• completely new doc design

Breaking changes

• Minuit uses its own, internal gradient by default. To change this back, use use_minuit_grad=False

• minimize(params=…) now filters correctly non-floating parameters.

• z.log has been moved to z.math.log (following TF)

Depreceations

Bug fixes and small changes

• ncalls is not correctly using the internal heuristc or the ncalls explicitly

• minimize(params=…) automatically extracts independent parameters.

• fix copy issue of KDEV1 and change name to ‘adaptive’ (instead of ‘adaptiveV1’)

• change exp name of lambda_ to lam (in init)

• add set_yield to BasePDF to allow setting the yield in place

• Fix possible bug in SumPDF with extended pdfs (automatically)

Experimental

Requirement changes

• upgrade to iminuit>=1.4

• remove cloudpickle hack fix

Thanks

Johannes for the docs re-design

0.5.3 (02.07.20)

Kernel density estimation for 1 dimension.

Major Features and Improvements

• add correlation method to FitResult

• Gaussian (Truncated) Kernel Density Estimation in one dimension zfit.pdf.GaussianKDE1DimV1 implementation with fixed and adaptive bandwidth added as V1. This is a feature that needs to be improved and feedback is welcome

• Non-relativistic Breit-Wigner PDF, called Cauchy, implementation added.

Breaking changes

• change human-readable name of Gauss, Uniform and TruncatedGauss to remove the ‘_tfp’ at the end of the name

Bug fixes and small changes

• fix color wrong in printout of results, params

• packaging: moved to pyproject.toml and a setup.cfg mainly, development requirements can be installed with the dev extra as (e.g.) pip install zfit[dev]

• Fix shape issue in TFP distributions for partial integration

• change zfit internal algorithm (zfit_error) to compute error/intervals from the profile likelihood, which is 2-3 times faster than previous algorithm.

• add from_minuit constructor to FitResult allowing to create it when using directly iminuit

• fix possible bias with sampling using accept-reject

Requirement changes

• pin down cloudpickle version (upstream bug with pip install) and TF, TFP versions

0.5.2 (13.05.2020)

Major Features and Improvements

• Python 3.8 and TF 2.2 support

• easier debugigng with set_graph_mode that can also be used temporarily with a context manager. False will make everything execute Numpy-like.

Bug fixes and small changes

• added get_params to loss

• fix a bug with the fixed_params when creating a sampler

• improve exponential PDF stability and shift when normalized

• improve accept reject sampling to account for low statistics

Requirement changes

• TensorFlow >= 2.2

0.5.1 (24.04.2020)

(0.5.0 was skipped)

Complete refactoring of Spaces to allow arbitrary function. New, more consistent behavior with extended PDFs. SumPDF refactoring, more explicit handling of fracs and yields. Improved graph building allowing for more fine-grained control of tracing. Stabilized minimization including a push-back for NaNs.

Major Features and Improvements

• Arbitrary limits as well as vectorization (experimental) are now fully supported. The new Space has an additional argument for a function that tests if a vector x is inside.

  To test if a value is inside a space, Space.inside can be used. To filter values, Space.filter.

  The limits returned are now by default numpy arrays with the shape (1, n_obs). This corresponds well to the old layout and can, using z.unstack_x(lower) be treated like Data. This has also some consequences for the output format of rect_area: this is now a vector.

  Due to the ambiguity of the name limits, area etc (since they do only reflect the rectangular case) method with leading rect_* have been added (rect_limits, rect_area etc.) and are encouraged to be used.

• Extending a PDF is more straightforward and removes any “magic”. The philosophy is: a PDF can be extended or not. But it does not change the fundamental behavior of functions.

• SumPDF has been refactored and behaves now as follows: Giving in pdfs (extended or not or mixed) and fracs (either length pdfs or one less) will create a non-extended SumPDF using the fracs. The fact that the pdfs are maybe extended is ignored. This will lead to highly consistent behavior. If the number of fracs given equals the number of pdfs, it is up to the user (currently) to take care of the normalization. Only if all pdfs are extended and no fracs are given, the sumpdf will be using the yields as normalized fracs and be extended.

• Improved graph building and z.function

  • the z.function can now, as with tf.function, be used either as a decorator without arguments or as a decorator with arguments. They are the same as in tf.function, except of a few additional ones.

  • zfit.run.set_mode allows to set the policy for whether everything is run in eager mode (graph=False), everything in graph, or most of it (graph=True) or an optimized variant, doing graph building only with losses but not just models (e.g. pdf won’t trigger a graph build, loss.value() will) with graph=’auto’.

  • The graph cache can be cleaned manually using zfit.run.clear_graph_cache() in order to prevent slowness in repeated tasks.

• Switch for numerical gradients has been added as well in zfit.run.set_mode(autograd=True/False).

• Resetting to the default can be done with zfit.run.set_mode_default()

• Improved stability of minimizer by adding penalty (currently in Minuit) as default. To have a better behavior with toys (e.g. never fail on NaNs but return an invalid FitResult), use the DefaultToyStrategy in zfit.mnimize.

• Exceptions are now publicly available in zfit.exception

• Added nice printout for FitResult and FitResult.params.

• get_params is now more meaningful, returning by default all independent parameters of the pdf, including yields. Arguments (floating, is_yield) allow for more fine-grained control.

Breaking changes

• Multiple limits are now handled by a MultiSpace class. Each Space has only “one limit” and no complicated layout has to be remembered. If you want to have a space that is defined in disconnected regions, use the + operator or functionally zfit.dimension.add_spaces

  To extract limits from multiple limits, MultiSpace and Space are both iterables, returning the containing spaces respectively itself (for the Space case).

• SumPDF changed in the behavior. Read above in the Major Features and Improvement.

• Integrals of extended PDFs are not extended anymore, but ext_integrate now returns the integral multiplied by the yield.

Deprecations

• ComposedParameter takes now params instead of dependents as argument, it acts now as the arguments to the value_fn. To stay future compatible, create e.g. def value_fn(p1, pa2) and using params = [‘param1, param2], value_fn will then be called as value_fn(param1, parma2). value_fn without arguments will probably break in the future.

• FitResult.error has been renamed to errors to better reflect that multiple errors, the lower and upper are returned.

Bug fixes and small changes

• fix a (nasty, rounding) bug in sampling with multiple limits

• fix bug in numerical calculation

• fix bug in SimplePDF

• fix wrong caching signature may lead to graph not being rebuild

• add zfit.param.set_values method that allows to set the values of multiple parameters with one command. Can, as the set_value method be used with a context manager.

• wrong size of weights when applying cuts in a dataset

• with_coords did drop axes/obs

• Fix function not traced when an error was raised during first trace

• MultipleLimits support for analytic integrals

• zfit.param.set_values(..) now also can use a FitResult as values argument to set the values from.

Experimental

• added a new error method, ‘zfit_error’ that is equivalent to ‘minuit_minos’, but not fully stable. It can be used with other minimizers as well, not only Minuit.

Requirement changes

• remove the outdated typing module

• add tableformatter, colored, colorama for colored table printout

Thanks

• Johannes Lade for code review and discussions.

• Hans Dembinski for useful inputs to the uncertainties.

0.4.3 (11.3.2020)

Major Features and Improvements

• refactor hesse_np with covariance matrix, make it available to all minimizers

Behavioral changes

Bug fixes and small changes

• fix bug in hesse_np

Requirement changes

Thanks

0.4.2 (27.2.2020)

Major Features and Improvements

• Refactoring of the Constraints, dividing into ProbabilityConstraint that can be sampled from and more general constraints (e.g. for parameter boundaries) that can not be sampled from.

• Doc improvements in the constraints.

• Add hesse error method (‘hesse_np’) available to all minimizers (not just Minuit).

Behavioral changes

• Changed default step size to an adaptive scheme, a fraction (1e-4) of the range between the lower and upper limits.

Bug fixes and small changes

• Add use_minuit_grad option to Minuit optimizer to use the internal gradient, often for more stable fits

• added experimental flag zfit.experimental_loss_penalty_nan, which adds a penalty to the loss in case the value is nan. Can help with the optimisation. Feedback welcome!

Requirement changes

Thanks

0.4.1 (12.1.20)

Release to keep up with TensorFlow 2.1

Major Features and Improvements

• Fixed the comparison in caching the graph (implementation detail) that leads to an error.

0.4.0 (7.1.2020)

This release switched to TensorFlow 2.0 eager mode. In case this breaks things for you and you need urgently a running version, install a version < 0.4.1. It is highly recommended to upgrade and make the small changes required.

Please read the upgrade guide <docs/project/upgrade_guide.rst> on a more detailed explanation how to upgrade.

TensorFlow 2.0 is eager executing and uses functions to abstract the performance critical parts away.

Major Features and Improvements

• Dependents (currently, and probably also in the future) need more manual tracking. This has mostly an effect on CompositeParameters and SimpleLoss, which now require to specify the dependents by giving the objects it depends (indirectly) on. For example, it is sufficient to give a ComplexParameter (which itself is not independent but has dependents) to a SimpleLoss as dependents (assuming the loss function depends on it).

• ComposedParameter does no longer allow to give a Tensor but requires a function that, when evaluated, returns the value. It depends on the dependents that are now required.

• Added numerical differentiation, which allows now to wrap any function with z.py_function (zfit.z). This can be switched on with zfit.settings.options[‘numerical_grad’] = True

• Added gradient and hessian calculation options to the loss. Support numerical calculation as well.

• Add caching system for graph to prevent recursive graph building

• changed backend name to z and can be used as zfit.z or imported from it. Added:

  • function decorator that can be used to trace a function. Respects dependencies of inputs and automatically caches/invalidates the graph and recreates.

  • py_function, same as tf.py_function, but checks and may extends in the future

  • math module that contains autodiff and numerical differentiation methods, both working with tensors.

Behavioral changes

• EDM goal of the minuit minimizer has been reduced by a factor of 10 to 10E-3 in agreement with the goal in RooFits Minuit minimizer. This can be varied by specifying the tolerance.

• known issue: the projection_pdf has troubles with the newest TF version and may not work properly (runs out of memory)

Bug fixes and small changes

Requirement changes

• added numdifftools (for numerical differentiation)

Thanks

0.3.7 (6.12.19)

This is a legacy release to add some fixes, next release is TF 2 eager mode only release.

Major Features and Improvements

• mostly TF 2.0 compatibility in graph mode, tests against 1.x and 2.x

Behavioral changes

Bug fixes and small changes

• get_depentents returns now an OrderedSet

• errordef is now a (hidden) attribute and can be changed

• fix bug in polynomials

Requirement changes

• added ordered-set

0.3.6 (12.10.19)

Special release for conda deployment and version fix (TF 2.0 is out)

This is the last release before breaking changes occur

Major Features and Improvements

• added ConstantParameter and zfit.param namespace

• Available on conda-forge

Behavioral changes

• an implicitly created parameter with a Python numerical (e.g. when instantiating a model) will be converted to a ConstantParameter instead of a fixed Parameter and therefore cannot be set to floating later on.

Bug fixes and small changes

• added native support TFP distributions for analytic sampling

• fix Gaussian (TFP Distribution) Constraint with mixed up order of parameters

• from_numpy automatically converts to default float regardless the original numpy dtype, dtype has to be used as an explicit argument

Requirement changes

• TensorFlow >= 1.14 is required

Thanks

• Chris Burr for the conda-forge deployment

0.3.4 (30-07-19)

This is the last release before breaking changes occur

Major Features and Improvements

• create Constraint class which allows for more fine grained control and information on the applied constraints.

• Added Polynomial models

• Improved and fixed sampling (can still be slightly biased)

Behavioral changes

None

Bug fixes and small changes

• fixed various small bugs

Thanks

for the contribution of the Constraints to Matthieu Marinangeli <[e-mail]>

0.3.3 (15-05-19)

Fixed Partial numeric integration

Bugfixes mostly, a few major fixes. Partial numeric integration works now.

Bugfixes

• data_range cuts are now applied correctly, also in several dimensions when a subset is selected (which happens internally of some Functors, e.g. ProductPDF). Before, only the selected obs was respected for cuts.

• parital integration had a wrong take on checking limits (now uses supports).

0.3.2 (01-05-19)

With 0.3.2, bugfixes and three changes in the API/behavior

Breaking changes

• tfp distributions wrapping is now different with dist_kwargs allowing for non-Parameter arguments (like other dists)

• sampling allows now for importance sampling (sampler in Model specified differently)

• model.sample now also returns a tensor, being consistent with pdf and integrate

Bugfixes

• shape handling of tfp dists was “wrong” (though not producing wrong results!), fixed. TFP distributions now get a tensor with shape (nevents, nobs) instead of a list of tensors with (nevents,)

Improvements

• refactor the sampling for more flexibility and performance (less graph constructed)

• allow to use more sophisticated importance sampling (e.g. phasespace)

• on-the-fly normalization (experimentally) implemented with correct gradient

0.3.1 (30-04-19)

Minor improvements and bugfixes including:

• improved importance sampling allowing to preinstantiate objects before it’s called inside the while loop

• fixing a problem with ztf.sqrt

0.3.0 (2019-03-20)

Beta stage and first pip release

0.0.1 (2018-03-22)

• First creation of the package.