Major Features and Improvements#

  • add JohnsonSU PDF, the Johnson SU distribution.

Breaking changes#


Bug fixes and small changes#


Requirement changes#


0.20.3 (19 Apr 2024)#

Bug fixes and small changes#

  • consistent behavior in loss: simple loss can take a gradient and hesse function and the default base loss provides fallbacks that work correctly between value_gradient and gradient. This maybe matters if you’ve implemented a custom loss and should fix any issues with it.

  • multiprocessing would get stuck due to an upstream bug in TensorFlow. Working around it by disabling an unused piece of code.


  • acampoverde for finding the bug in the multiprocessing

0.20.2 (16 Apr 2024)#

Two small bugfixes - fix backwards incompatible change of sampler - detect if a RegularBinning has been transformed, raise error.

0.20.1 (14 Apr 2024)#

Major Features and Improvements#

  • fix dumping and add convenience wrapper zfit.dill to dump and load objects with dill (a more powerful pickle). This way, any zfit object can be saved and loaded, such as FitResult that contains all other important objects to recreate the fit.

  • improved performance for numerical gradient calculation, fixing also a minor numerical issue.

Bug fixes and small changes#

  • runing binned fits without a graph could deadlock, fixed.

0.20.0 (12 Apr 2024)#

Complete overhaul of zfit with a focus on usability and a variety of new pdfs!

Major Features and Improvements#

  • Parameter behavior has changed, multiple parameters with the same name can now coexist! The NameAlreadyTakenError has been successfully removed (yay!). The new behavior only enforces that names and matching parameters within a function/PDF/loss are unique, as otherwise inconsistent expectations appear (for the full discussion on this, see here).

  • Space and limits have a complete overhaul in front of them, in short, these overcomplicated objects get simplified and the limits become more usable, in terms of dimensions. The full discussion and changes can be found here .

  • add an unbinned Sampler to the public namespace under this object is returned in the create_sampler method and allows to resample from a function without recreating the compiled function, i.e. loss. It has an additional method update_data to update the data without recompiling the loss and can be created from a sample only. Useful to have a custom dataset in toys.

  • allow to use pandas DataFrame as input where zfit Data objects are expected

  • Methods of PDFs and loss functions that depend on parameters take now the value of a parameter explicitly as arguments, as a mapping of str (parameter name) to value.

  • Python 3.12 support

  • add GeneralizedCB PDF which is similar to the DoubleCB PDF but with different standard deviations for the left and right side.

  • Added functor for PDF caching CachedPDF: pdf, integrate PDF methods can be cacheable now

  • add faddeeva_humlicek function under the zfit.z.numpy namespace. This is an implementation of the Faddeeva function, combining Humlicek’s rational approximations according to Humlicek (JQSRT, 1979) and Humlicek (JQSRT, 1982).

  • add Voigt profile PDF which is a convolution of a Gaussian and a Cauchy distribution.

  • add TruncatedPDF that allows to truncate in one or multiple ranges (replaces “MultipleLimits” and “MultiSpace”)

  • add LogNormal PDF, a log-normal distribution, which is a normal distribution of the logarithm of the variable.

  • add ChiSquared PDF, the standard chi2 distribution, taken from tensorflow-probability implementation.

  • add StudentT PDF, the standard Student’s t distribution, taken from tensorflow-probability implementation.

  • add GaussExpTail and GeneralizedGaussExpTail PDFs, which are a Gaussian with an exponential tail on one side and a Gaussian with different sigmas on each side and different exponential tails on each side respectively.

  • add QGauss PDF, a distribution that arises from the maximization of the Tsallis entropy under appropriate constraints, see here.

  • add BifurGauss PDF, a Gaussian distribution with different sigmas on each side of the mean.

  • add Bernstein PDF, which is a PDF defined by a linear combination of Bernstein polynomials given their coefficients.

  • add Gamma PDF, the Gamma distribution.

  • Data has now a with_weights method that returns a new data object with different weights and an improved with_obs that allows to set obs with new limits. These replace the set_weights and set_data_range methods for a more functional approach.

  • add label to different objects (PDF, Data, etc.) that allows to give a human-readable name to the object. This is used in the plotting and can be used to identify objects. Notably, Parameters have a label that can be arbitrary. Space has one label for each observable if the space is a product of spaces. Space.label is a string and only possible for one-dimensional spaces, while Space.labels is a list of strings and can be used for any, one- or multi-dimensional spaces.

  • add to concatenate multiple data objects into one along the index or along the observables. Similar to pd.concat.

  • PDFs now have a to_truncated method that allows to create a truncated version of the PDF, possibly with different and multiple limits. This allows to easily create a PDF with disjoint limits.

  • Data and PDF that take obs in the initialization can now also take binned observables, i.e. a zfit.Space with binning=... and will return a binned version of the object ( or zfit.pdf.BinnedFromUnbinned, where the latter is a generic wrapper). This is equivalent of calling to_binned on the objects)

  • zfit.Data can be instantiated directly with most data types, such as numpy arrays, pandas DataFrames etc insead of using the dedicated constructors from_numpy, from_pandas etc. The constructors may still provide additional functionality, but overall, the switch should be seamless.

Breaking changes#

This release contains multiple “breaking changes”, however, the vast majority if not all apply only for edge cases and undocummented functions.

  • a few arguments are now keyword-only arguments. This can break existing code if the arguments were given as positional arguments. Just use the appropriate keyword arguments instead. (Example: instead of using zfit.Space(obs, limits) use zfit.Space(obs, limits=limits)). This was introduced to make the API more robust and to avoid errors due to the order of arguments, with a few new ways of creating objects.

  • Data.from_root: deprecated arguments branches and branch_aliases have been removed. Use obs and obs_aliases instead.

  • NameAlreadyTakenError was removed, see above for the new behavior. This should not have an effect on any existing code except if you relied on the error being thrown.

  • Data objects had an intrinsic, TensorFlow V1 legacy behavior: they were actually cut when the data was retrieved. This is now changed and the data is cut when it is created. This should not have any impact on existing code and just improve runtime and memory usage.

  • Partial integration used to use some broadcasting tricks that could potentially fail. It uses now a dynamic while loop that _could_ be slower but works for arbitrary PDFs. This should not have any impact on existing code and just improve stability (but technically, the data given to the PDF if doing partial integration is now “different”, in the sense that it’s now not different anymore from any other call)

  • if a tf.Variable was used to store the number of sampled values in a sampler, it was possible to change the value of that variable to change the number of samples drawn. This is now not possible anymore and the number of samples should be given as an argument n to the resample method, as was possible since a long time.

  • create_sampler has a breaking change for fixed_params: when the argument was set to False, any change in the parameters would be reflected when resampling. This highly statebased behavior was confusing and is now removed. The argument is now called params and behaves as expected: the sampler will remember the parameters at the time of creation, possibly updated with params and will not change anymore. To sample from a different set of parameters, the params have to be passed to the resample method _explicitly_.

  • the default names for hesse and errors have now been changed to hesse and errors, respectively. This was deprecated since a while and both names were available for backwards compatibility. The old names are now removed. If you get an error, minuit_hessse or minuit_minos not found, just replace it with hesse and errors.


  • result.fminfull is deprecated and will be removed in the future. Use result.fmin instead.

  • Data.set_data_range is deprecated and will be removed in the future. Use with_range instead.

  • Space has many deprecated methods, such as rect_limits and quite a few more. The full discussion can be found here.

  • fixed_params in create_sampler is deprecated and will be removed in the future. Use params instead.

  • fixed_params attribute of the Sampler is deprecated and will be removed in the future. Use params instead.

  • uncertainties in GaussianConstraint is deprecated and will be removed in the future. Use either explicitly sigma or cov.

  • the ComposedParameter and ComplexParameter argument value_fn is deprecated in favor of the new argument func. Identical behavior.

  • is deprecated and will be removed in the future. Simply remove it should work in most cases. (if an explicity numpy, not just array-like, cast is needed, use np.asarray(...). But usually this is not needed). This function is an old relic from the past TensorFlow 1.x, tf.Session times and is not needed anymore. We all remember well these days :)

Bug fixes and small changes#

  • complete overhaul of partial integration that used some broadcasting tricks that could potentially fail. It uses now a dynamic while loop that _could_ be slower but works for arbitrary PDFs and no problems should be encountered anymore.

  • FitResult can now be used as a context manager, which will automatically set the values of the parameters to the best fit values and reset them to the original values after the context is left. A new method update_params allows to update the parameters with the best fit values explicitly.

  • result.fmin now returns the full likelihood, while result.fminopt returns the optimized likelihood with potential constant subtraction. The latter is mostly used by the minimizer and other libraries. This behavior is consistent with the behavior of other methods in the loss that return by default the full, unoptimized value.

  • serialization only allowed for one specific limit (space) of each obs. Multiple, independent limits can now be serialized.

  • Increased numerical stability: this was compromised due to some involuntary float32 conversions in TF. This has been fixed.

  • arguments sigma and cov are now used in GaussianConstraint, both mutually exclusive, to ensure the intent is clear.

  • improved hashing and precompilation in loss, works now safely also with samplers.

  • seed setting is by default completely randomized. This is a change from the previous behavior where the seed was set to a more deterministic value. Use seeds only for reproducibility and not for real randomness, as some strange correlations between seeds have been observed. To guarantee full randomness, just call without arguments.

  • now returns the seed that was set. This is useful for reproducibility.


  • a simple plot mechanism has been added with pdf.plot.plotpdf to plot PDFs. This is simple and fully interacts with matplotlib, allowing to plot quickly in a more interactive way.

  • this is an experimental feature that allows to disable the parameter update in a fit as is currently done whenever minimize is called. In conjunction with the new method update_params(), this can be used as result = minimizer.minimize(...).update_params() to keep the same behavior as currently. Also, the context manager of FitResult can be used to achieve the same behavior in a context manager (with minimizer.minimize(…) as result: …) also works.

Requirement changes#

  • upgrade to TensorFlow 2.16 and TensorFlow Probability 0.24


  • huge thanks to @ikrommyd (Iason Krommydas) for the addition of various PDFs and to welcome him on board as a new contributor!

  • @anjabeck for the addition of the ChiSquared PDF

0.18.2 (13 Mar 2024)#

Hotfix for missing dependency attrs

0.18.1 (22 Feb 2024)#

Bug fixes in randomness and improved caching

Major Features and Improvements#

  • reduced the number of graph caching reset, resulting in significant speedups in some cases

Bug fixes and small changes#

  • use random generated seeds for numpy and TF, as they can otherwise have unwanted correlations


@anjabeck for the bug report and the idea to use random seeds for numpy and TF @acampoverde for reporting the caching issue

0.18.0 (29 Jan 2024)#

Major Features and Improvements#

  • update to TensorFlow 2.15, TensorFlow Probability 0.23

  • drop Python 3.8 support

0.17.0 (29 Jan 2024)#

TensorFlow 2.15, drop Python 3.8 support

Major Features and Improvements#

  • add correct uncertainty for unbinned, weighted fits with constraints and/or that are extended.

  • allow mapping in zfit.param.set_values for values

Bug fixes and small changes#

  • fix issues where EDM goes negative, set to 999

  • improved stability of the loss evaluation

  • improved uncertainty calculation accuracy with zfit_error


Daniel Craik for the idea of allowing a mapping in set_values

0.16.0 (26 July 2023)#

Major Features and Improvements#

  • add full option to loss call of value, which returns the unoptimized value allowing for easier statistical tests using the loss. This is the default behavior and should not break any backwards compatibility, as the “not full loss” was arbitrary.

  • changed the FitResult to print both loss values, the unoptimized (full) and optimized (internal)

Bug fixes and small changes#

  • bandwidth preprocessing was ignored in KDE

  • unstack_x with an obs as argument did return the wrong shape


@schmitse for reporting the bug in the KDE bandwidth preprocessing @lorenzopaolucci for bringing up the absolute value of the loss in the fitresult as an issue

0.15.5 (26 July 2023)#

Bug fixes and small changes#

  • fix a bug in histmodifier that would not properly take into account the yield of the wrapped PDF

0.15.2 (20 July 2023)#

Fix missing attrs dependency

Major Features and Improvements#

  • add option full in loss to return the full, unoptimized value (currently not default), allowing for easier statistical tests using the loss

0.15.0 (13 July 2023)#

Update to TensorFlow 2.13.x

Requirement changes#

  • TensorFlow upgraded to ~=2.13.0

  • as TF 2.13.0 ships with the arm64 macos wheels, the requirement of tensorflow_macos is removed


  • Iason Krommydas for helping with the macos requirements for TF

0.14.1 (1 July 2023)#

Major Features and Improvements#

  • zfit broke for pydantic 2, which upgraded.

Requirement changes#

  • restrict pydantic to <2.0.0

0.14.0 (22 June 2023)#

Major Features and Improvements#

  • support for Python 3.11, dropped support for Python 3.7

Bug fixes and small changes#

-fix longstanding bug in parameters caching

Requirement changes#

  • update to TensorFlow 2.12

  • removed tf_quant_finance

0.13.2 (15. June 2023)#

Bug fixes and small changes#

  • fix a caching problem with parameters (could cause issues with larger PDFs as params would be “remembered” wrongly)

  • more helpful error message when jacobian (as used for weighted corrections) is analytically asked but fails

  • make analytical gradient for CB integral work

0.13.1 (20 Apr 2023)#

Bug fixes and small changes#

  • array bandwidth for KDE works now correctly

Requirement changes#

  • fixed uproot for Python 3.7 to <5


  • @schmitse for reporting and solving the bug in the KDE bandwidth with arrays

0.13.0 (19 April 2023)#

Major Features and Improvements#

last Python 3.7 version

Bug fixes and small changes#

  • SampleData is not used anymore, a Data object is returned (for simple sampling). The create_sampler will still return a SamplerData object though as this differs from Data.


  • Added support on a best-effort for human-readable serialization of objects including an HS3-like representation, find a tutorial on serialization here. Most built-in unbinned PDFs are supported. This is still experimental and not yet fully supported. Dumping can be performed safely, loading maybe easily breaks (also between versions), so do not rely on it yet. Everything else - apart of trying to dump - should only be used for playing around and giving feedback purposes.

Requirement changes#

  • allow uproot 5 (remove previous restriction)


  • to Johannes Lade for the amazing work on the serialization, which made this HS3 implementation possible!

0.12.1 (1 April 2023)#

Bug fixes and small changes#

  • added extended as a parameter to all PDFs: a PDF can now directly be extended without the need for create_extended (or set_yield).

  • to_pandas and from_pandas now also support weights as columns. Default column name is "".

  • add numpy and backend to options when setting the seed

  • reproducibility by fixing the seed in zfit is restored, now also sets the seed for the backend(numpy, tensorflow, etc.) if requested (on by default)


  • Sebastian Schmitt @schmitse for reporting the bug in the non-reproducibility of the seed.

0.12.0 (13 March 2023)#

Bug fixes and small changes#

  • create_extended added None to the name, removed.

  • SimpleConstraint now also takes a function that has an explicit params argument.

  • add name argument to create_extended.

  • adding binned losses would error due to the removed fit_range argument.

  • setting a global seed made the sampler return constant values, fixed (unoptimized but correct). If you ran a fit with a global seed, you might want to rerun it.

  • histogramming and limit checks failed due to a stricter Numpy check, fixed.


  • @P-H-Wagner for finding the bug in SimpleConstraint.

  • Dan Johnson for finding the bug in the binned loss that would fail to sum them up.

  • Hanae Tilquin for spotting the bug with TensorFlows changed behavior or random states inside a tf.function, leading to biased samples whenever a global seed was set.

0.11.1 (20 Nov 2022)#

Hotfix for wrong import

0.11.0 (29 Nov 2022)#

Major Features and Improvements#

  • columns of unbinned data can be accessed with the obs like a mapping (like a dataframe)

  • speedup builtin errors method and make it more robust

Breaking changes#

  • Data can no longer be used directly as an array-like object but got mapping-like behavior.

  • some old deprecated methods were removed

Bug fixes and small changes#

  • improved caching speed, reduced tradeoff against memory

  • yields were not added correctly in some (especially binned) PDFs and the fit would fail

Requirement changes#

  • add jacobi (many thanks at @HansDembinski for the package)

0.10.1 (31 Aug 2022)#

Major Features and Improvements#

  • reduce the memory footprint on (some) fits, especially repetitive (loops) ones. Reduces the number of cached compiled functions. The cachesize can be set with and specifies the number of compiled functions that are kept in memory. The default is 10, but this can be tuned. Lower values can reduce memory usage, but potentially increase runtime.

Bug fixes and small changes#

  • Enable uniform binning for n-dimensional distributions with integer(s).

  • Sum of histograms failed for calling the pdf method (can be indirectly), integrated over wrong axis.

  • Binned PDFs expected binned spaces for limits, now unbinned limits are also allowed and automatically

    converted to binned limits using the PDFs binning.

  • Speedup sampling of binned distributions.

  • add to_binned and to_unbinned methods to PDF


  • Justin Skorupa for finding the bug in the sum of histograms and the missing automatic conversion of unbinned spaces to binned spaces.

0.10.0 (22. August 2022)#

Public release of binned fits and upgrade to Python 3.10 and TensorFlow 2.9.

Major Features and Improvements#

  • improved data handling in constructors from_pandas (which allows now to have weights as columns, dataframes that are a superset of the obs) and from_root (obs can now be spaces and therefore cuts can be direcly applied)

  • add hashing of unbinned datasets with a hashint attribute. None if no hash was possible.

Breaking changes#


Bug fixes and small changes#

  • SimpleLoss correctly supports both functions with implicit and explicit parameters, also if they are decorated.

  • extended sampling errored for some cases of binned PDFs.

  • ConstantParameter errored when converted to numpy.

  • Simultaneous binned fits could error with different binning due to a missing sum over a dimension.

  • improved stability in loss evaluation of constraints and poisson/chi2 loss.

  • reduce gradient evaluation time in errors for many parameters.

  • Speedup Parameter value assignement in fits, which is most notably when the parameter update time is comparably large to the fit evaluation time, such as is the case for binned fits with many nuisance parameters.

  • fix ipyopt was not pickleable in a fitresult

  • treat parameters sometimes as “stateless”, possibly reducing the number of retraces and reducing the memory footprint.


Requirement changes#

  • nlopt and ipyopt are now optional dependencies.

  • Python 3.10 added

  • TensorFlow >= 2.9.0, <2.11 is now required and the corresponding TensorFlow-Probability version >= 0.17.0, <0.19.0


  • @YaniBion for discovering the bug in the extended sampling and testing the alpha release

  • @ResStump for reporting the bug with the simultaneous binned fit


Major Features and Improvements#

  • Save results by pickling, unpickling a frozen (FitResult.freeze()) result and using zfit.param.set_values(params, result) to set the values of params.


  • the default name of the uncertainty methods hesse and errors depended on the method used (such as "minuit_hesse", "zfit_errors" etc.) and would be the exact method name. New names are now ‘hesse’ and ‘errors’, independent of the method used. This reflects better that the methods, while internally different, produce the same result. To update, use ‘hesse’ instead of ‘minuit_hesse’ or ‘hesse_np’ and ‘errors’ instead of ‘zfit_errors’ or "minuit_minos" in order to access the uncertainties in the fitresult. Currently, the old names are still available for backwards compatibility. If a name was explicitly chosen in the error method, nothing changed.

Bug fixes and small changes#

  • KDE datasets are now correctly mirrored around observable space limits

  • multinomial sampling would return wrong results when invoked multiple times in graph mode due to a non-dynamic shape. This is fixed and the sampling is now working as expected.

  • increase precision in FitResult string representation and add that the value is rounded


  • schmitse for finding and fixing a mirroring bug in the KDEs

  • Sebastian Bysiak for finding a bug in the multinomial sampling


Major Features and Improvements#

  • Binned fits support, although limited in content, is here! This includes BinnedData, binned PDFs, and binned losses. TODO: extend to include changes/point to binned introduction.

  • new Poisson PDF

  • added Poisson constraint, LogNormal Constraint

  • Save results by pickling, unpickling a frozen (FitResult.freeze()) result and using zfit.param.set_values(params, result) to set the values of params.

Breaking changes#

  • params given in ComposedParameters are not sorted anymore. Rely on their name instead.

  • norm_range is now called norm and should be replaced everywhere if possible. This will break in the future.


Bug fixes and small changes#

  • remove warning when using rect_limits or similar.

  • gauss integral accepts now also tensor inputs in limits

  • parameters at limits is now shown correctly


Requirement changes#

  • add TensorFlow 2.7 support


0.8.3 (5 Apr 2022)#

  • fixate nlopt to < 2.7.1

0.8.2 (20 Sep 2021)#

Bug fixes and small changes#

  • fixed a longstanding bug in the DoubleCB implementation of the integral.

  • remove outdated deprecations

0.8.1 (14. Sep. 2021)#

Major Features and Improvements#

  • allow FitResult to freeze(), making it pickleable. The parameters are replaced by their name, the objects such as loss and minimizer as well.

  • improve the numerical integration by adding a one dimensional efficient integrator, testing for the accuracy of multidimensional integrals. If there is a sharp peak, this maybe fails to integrate and the number of points has to be manually raised

  • add highly performant kernel density estimation (mainly contributed by Marc Steiner) in 1 dimension which allow for the choice of arbitrary kernels, support boundary mirroring of the data and allow for large (millions) of data samples: - KDE1DimExact for the normal density estimation - KDE1DimGrid using a binning - KDE1DimFFT using a binning and FFT - KDE1DimISJ using a binning and an algorithm (ISJ) to solve the optimal bandwidth

    For an introduction, see either Kernel Density Estimation or the tutorial Model

  • add windows in CI

Breaking changes#

  • the numerical integration improved with more sensible values for tolerance. This means however that some fits will greatly increase the runtime. To restore the old behavior globally, do for each instance pdf.update_integration_options(draws_per_dim=40_000, max_draws=40_000, tol=1) This will integrate regardless of the chosen precision and it may be non-optimal. However, the precision estimate in the integrator is also not perfect and maybe overestimates the error, so that the integration by default takes longer than necessary. Feel free to play around with the parameters and report back.

Bug fixes and small changes#

  • Double crystallball: move a minus sign down, vectorize the integral, fix wrong output shape of pdf

  • add a minimal value in the loss to avoid NaNs when taking the log of 0

  • improve feedback when taking the derivative with respect to a parameter that a function does not depend on or if the function is purely Python.

  • make parameters deletable, especially it works now to create parameters in a function only and no NameAlreadyTakenError will be thrown.

Requirement changes#

  • add TensorFlow 2.6 support (now 2.5 and 2.6 are supported)


  • Marc Steiner for contributing many new KDE methods!

0.7.2 (7. July 2021)#

Bug fixes and small changes#

  • fix wrong arguments to minimize

  • make BaseMinimizer arguments optional

0.7.1 (6. July 2021)#

Bug fixes and small changes#

  • make loss callable with array arguments and therefore combatible with iminuit cost functions.

  • fix a bug that allowed FitResults to be valid that are actually invalid (reported by Maxime Schubiger).

0.7.0 (03 Jun 2021)#

Major Features and Improvements#

  • add Python 3.9 support

  • upgrade to TensorFlow 2.5

Bug fixes and small changes#

  • Scipy minimizers with hessian arguments use now BFGS as default

Requirement changes#

  • remove Python 3.6 support

  • boost-histogram

0.6.6 (12.05.2021)#

Update ipyopt requirement < 0.12 to allow numpy compatible with TensorFlow

0.6.5 (04.05.2021)#

  • hotfix for wrong argument in exponential PDF

  • removed requirement ipyopt, can be installed with pip install zfit[ipyopt] or by manually installing pip install ipyopt

0.6.4 (16.4.2021)#

Bug fixes and small changes#

  • remove requirement of Ipyopt on MacOS as no wheels are available. This rendered zfit basically non-installable.

0.6.3 (15.4.2021)#

Bug fixes and small changes#

  • fix loss failed for large datasets

  • catch hesse failing for iminuit


Minor small fixes.

Bug fixes and small changes#

  • add loss to callback signature that gives full access to the model

  • add create_new() to losses in order to re-instantiate them with new models and data preserving their current (and future) options and other arguments

0.6.1 (31.03.2021)#

Release for fix of minimizers that performed too bad

Breaking changes#

  • remove badly performing Scipy minimizers ScipyTrustKrylovV1 and ScipyTrustNCGV1

Bug fixes and small changes#

  • fix auto conversion to complex parameter using constructor

0.6.0 (30.3.2021)#

Added many new minimizers from different libraries, all with uncertainty estimation available.

Major Features and Improvements#

  • upgraded to TensorFlow 2.4

  • Added many new minimizers. A full list can be found in Minimize.

  • Completely new and overhauled minimizers design, including:

    • minimizers can now be used with arbitrary Python functions and an initial array independent of zfit

    • a minimization can be ‘continued’ by passing init to minimize

    • more streamlined arguments for minimizers, harmonized names and behavior.

    • Adding a flexible criterion (currently EDM) that will terminate the minimization.

    • Making the minimizer fully stateless.

    • Moving the loss evaluation and strategy into a LossEval that simplifies the handling of printing and NaNs.

    • Callbacks are added to the strategy.

  • Major overhaul of the FitResult, including:

    • improved zfit_error (equivalent of MINOS)

    • minuit_hesse and minuit_minos are now available with all minimizers as well thanks to an great improvement in iminuit.

    • Added an approx hesse that returns the approximate hessian (if available, otherwise empty)

  • upgrade to iminuit v2 changes the way it works and also the Minuit minimizer in zfit, including a new step size heuristic. Possible problems can be caused by iminuit itself, please report in case your fits don’t converge anymore.

  • improved compute_errors in speed by caching values and the reliability by making the solution unique.

  • increased stability for large datasets with a constant subtraction in the NLL

Breaking changes#

  • NLL (and extended) subtracts now by default a constant value. This can be changed with a new options argument. COMPARISON OF DIFFEREN NLLs (their absolute values) fails now! (flag can be deactivated)

  • BFGS (from TensorFlow Probability) has been removed as it is not working properly. There are many alternatives such as ScipyLBFGSV1 or NLoptLBFGSV1

  • Scipy (the minimizer) has been removed. Use specialized Scipy* minimizers instead.

  • Creating a zfit.Parameter, usign set_value or set_values now raises a ValueError if the value is outside the limits. Use assign to suppress it.


  • strategy to minimizer should now be a class, not an instance anymore.

Bug fixes and small changes#

  • zfit_error moved only one parameter to the correct initial position. Speedup and more reliable.

  • FFTconv was shifted if the kernel limits were not symetrical, now properly taken into account.

  • circumvent overflow error in sampling

  • shuffle samples from sum pdfs to ensure uniformity and remove conv sampling bias

  • create_sampler now samples immediately to allow for precompile, a new hook that will allow objects to optimize themselves.

Requirement changes#

  • ipyopt

  • nlopt

  • iminuit>=2.3

  • tensorflow ~= 2.4

  • tensorflow-probability~=12

For devs: - pre-commit - pyyaml - docformatter


  • Hans Dembinski for the help on upgrade to imituit V2

  • Thibaud Humair for helpful remarks on the parameters

0.5.6 (26.1.2020)#

Update to fix iminuit version

Bug fixes and small changes#

  • Fix issue when using a ComposedParameter as the rate argument of a Poisson PDF

Requirement changes#

  • require iminuit < 2 to avoid breaking changes

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

Breaking changes#


Bug fixes and small changes#

  • print parameter inside function context works now correctly


  • 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


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)

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)


Requirement changes#

  • upgrade to iminuit>=1.4

  • remove cloudpickle hack fix


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.

    • 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 in order to prevent slowness in repeated tasks.

  • Switch for numerical gradients has been added as well in

  • Resetting to the default can be done with

  • 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.


  • 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.


  • 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


  • 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#


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#


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)


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


  • 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#


Bug fixes and small changes#

  • fixed various small bugs


for the contribution of the Constraints to Matthieu Marinangeli <>

0.3.3 (15-05-19)#

Fixed Partial numeric integration

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

  • 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


  • 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,)


  • 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.