fullargument for binned NLLs was not working properly and return a partially optimized loss value.- jit compile all methods of the loss (gradient, hessian) to avoid recompilation every time. This can possibly speed up different minimizers significantly.
- add
JohnsonSUPDF, the Johnson SU distribution.
- increase reliability of
zfit.dill.dumpandzfit.dill.dumpswith an additionalverifyargument that reloads the dumped object to verify it was correctly dumped and retries if it wasn't. - fix missing imported namespaces
- fixed a memory leak when creating multiple parameters
- add data loaders to
zfit.datanamespace
- upgrade to TensorFlow 2.17 and TensorFlow Probability 0.25
- Davide Lancierini for finding and helping to debug the dill dumping issue
- James Herd for finding and reproducing the memory leak
- 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_gradientandgradient. 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
Two small bugfixes - fix backwards incompatible change of sampler - detect if a RegularBinning has been transformed, raise error.
- fix dumping and add convenience wrapper
zfit.dillto dump and load objects with dill (a more powerful pickle). This way, any zfit object can be saved and loaded, such asFitResultthat contains all other important objects to recreate the fit. - improved performance for numerical gradient calculation, fixing also a minor numerical issue.
- runing binned fits without a graph could deadlock, fixed.
Complete overhaul of zfit with a focus on usability and a variety of new pdfs!
- Parameter behavior has changed, multiple parameters with the same name can now coexist!
The
NameAlreadyTakenErrorhas 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). Spaceand 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
Samplerto the public namespace underzfit.data.Sampler: this object is returned in thecreate_samplermethod and allows to resample from a function without recreating the compiled function, i.e. loss. It has an additional methodupdate_datato 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
GeneralizedCBPDF which is similar to theDoubleCBPDF but with different standard deviations for the left and right side. - Added functor for PDF caching
CachedPDF:pdf,integratePDF methods can be cacheable now - add
faddeeva_humlicekfunction under thezfit.z.numpynamespace. This is an implementation of the Faddeeva function, combining Humlicek's rational approximations according to Humlicek (JQSRT, 1979) and Humlicek (JQSRT, 1982). - add
Voigtprofile PDF which is a convolution of a Gaussian and a Cauchy distribution. - add
TruncatedPDFthat allows to truncate in one or multiple ranges (replaces "MultipleLimits" and "MultiSpace") - add
LogNormalPDF, a log-normal distribution, which is a normal distribution of the logarithm of the variable. - add
ChiSquaredPDF, the standard chi2 distribution, taken from tensorflow-probability implementation. - add
StudentTPDF, the standard Student's t distribution, taken from tensorflow-probability implementation. - add
GaussExpTailandGeneralizedGaussExpTailPDFs, 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
QGaussPDF, a distribution that arises from the maximization of the Tsallis entropy under appropriate constraints, see here. - add
BifurGaussPDF, a Gaussian distribution with different sigmas on each side of the mean. - add
BernsteinPDF, which is a PDF defined by a linear combination of Bernstein polynomials given their coefficients. - add
GammaPDF, the Gamma distribution. Datahas now awith_weightsmethod that returns a new data object with different weights and an improvedwith_obsthat allows to set obs with new limits. These replace theset_weightsandset_data_rangemethods for a more functional approach.- add
labelto 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.Spacehas one label for each observable if the space is a product of spaces.Space.labelis a string and only possible for one-dimensional spaces, whileSpace.labelsis a list of strings and can be used for any, one- or multi-dimensional spaces. - add
zfit.data.concat(...)to concatenate multiple data objects into one along the index or along the observables. Similar topd.concat. - PDFs now have a
to_truncatedmethod 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. DataandPDFthat takeobsin the initialization can now also take binned observables, i.e. azfit.Spacewithbinning=...and will return a binned version of the object (zfit.data.BinnedDataorzfit.pdf.BinnedFromUnbinned, where the latter is a generic wrapper). This is equivalent of callingto_binnedon the objects)zfit.Datacan be instantiated directly with most data types, such as numpy arrays, pandas DataFrames etc insead of using the dedicated constructorsfrom_numpy,from_pandasetc. The constructors may still provide additional functionality, but overall, the switch should be seamless.
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)usezfit.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 argumentsbranchesandbranch_aliaseshave been removed. Useobsandobs_aliasesinstead.NameAlreadyTakenErrorwas 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.Variablewas 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 argumentnto theresamplemethod, as was possible since a long time. create_samplerhas a breaking change forfixed_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 calledparamsand behaves as expected: the sampler will remember the parameters at the time of creation, possibly updated withparamsand will not change anymore. To sample from a different set of parameters, the params have to be passed to theresamplemethod _explicitly_.- the default names for
hesseanderrorshave now been changed tohesseanderrors, 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_hessseorminuit_minosnot found, just replace it withhesseanderrors.
result.fminfullis deprecated and will be removed in the future. Useresult.fmininstead.Data.set_data_rangeis deprecated and will be removed in the future. Usewith_rangeinstead.Spacehas many deprecated methods, such asrect_limitsand quite a few more. The full discussion can be found here.fixed_paramsincreate_sampleris deprecated and will be removed in the future. Useparamsinstead.fixed_paramsattribute of theSampleris deprecated and will be removed in the future. Useparamsinstead.uncertaintiesinGaussianConstraintis deprecated and will be removed in the future. Use either explicitlysigmaorcov.- the
ComposedParameterandComplexParameterargumentvalue_fnis deprecated in favor of the new argumentfunc. Identical behavior. zfit.run(...)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, usenp.asarray(...). But usually this is not needed). This function is an old relic from the past TensorFlow 1.x,tf.Sessiontimes and is not needed anymore. We all remember well these days :)
- 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.
FitResultcan 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 methodupdate_paramsallows to update the parameters with the best fit values explicitly.result.fminnow returns the full likelihood, whileresult.fminoptreturns 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
sigmaandcovare now used inGaussianConstraint, 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
zfit.run.set_seed()without arguments. zfit.run.set_seednow returns the seed that was set. This is useful for reproducibility.
- a simple
plotmechanism has been added withpdf.plot.plotpdfto plot PDFs. This is simple and fully interacts with matplotlib, allowing to plot quickly in a more interactive way. zfit.run.experimental_disable_param_update: this is an experimental feature that allows to disable the parameter update in a fit as is currently done wheneverminimizeis called. In conjunction with the new methodupdate_params(), this can be used asresult = minimizer.minimize(...).update_params()to keep the same behavior as currently. Also, the context manager ofFitResultcan be used to achieve the same behavior in a context manager (with minimizer.minimize(...) as result: ...) also works.
- 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
ChiSquaredPDF
Hotfix for missing dependency attrs
Bug fixes in randomness and improved caching
- reduced the number of graph caching reset, resulting in significant speedups in some cases
- 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
- update to TensorFlow 2.15, TensorFlow Probability 0.23
- drop Python 3.8 support
TensorFlow 2.15, drop Python 3.8 support
- add correct uncertainty for unbinned, weighted fits with constraints and/or that are extended.
- allow mapping in
zfit.param.set_valuesfor values
- 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
- add
fulloption to loss call ofvalue, 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
FitResultto print both loss values, the unoptimized (full) and optimized (internal)
- bandwidth preprocessing was ignored in KDE
unstack_xwith anobsas 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
- fix a bug in histmodifier that would not properly take into account the yield of the wrapped PDF
Fix missing attrs dependency
- add option
fullin loss to return the full, unoptimized value (currently not default), allowing for easier statistical tests using the loss
Update to TensorFlow 2.13.x
- TensorFlow upgraded to ~=2.13.0
- as TF 2.13.0 ships with the arm64 macos wheels, the requirement of
tensorflow_macosis removed
- Iason Krommydas for helping with the macos requirements for TF
- zfit broke for pydantic 2, which upgraded.
- restrict pydantic to <2.0.0
- support for Python 3.11, dropped support for Python 3.7
-fix longstanding bug in parameters caching
- update to TensorFlow 2.12
- removed
tf_quant_finance
- 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
- array bandwidth for KDE works now correctly
- fixed uproot for Python 3.7 to <5
- @schmitse for reporting and solving the bug in the KDE bandwidth with arrays
last Python 3.7 version
SampleDatais not used anymore, aDataobject is returned (for simple sampling). Thecreate_samplerwill still return aSamplerDataobject though as this differs fromData.
- 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.
- allow uproot 5 (remove previous restriction)
- to Johannes Lade for the amazing work on the serialization, which made this HS3 implementation possible!
- added
extendedas a parameter to all PDFs: a PDF can now directly be extended without the need forcreate_extended(orset_yield). to_pandasandfrom_pandasnow also support weights as columns. Default column name is"".- add
numpyandbackendto options when setting the seed - reproducibility by fixing the seed in zfit is restored,
zfit.run.set_seednow 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.
create_extendedaddedNoneto the name, removed.SimpleConstraintnow also takes a function that has an explicitparamsargument.- add
nameargument tocreate_extended. - adding binned losses would error due to the removed
fit_rangeargument. - 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.
Hotfix for wrong import
- columns of unbinned
datacan be accessed with the obs like a mapping (like a dataframe) - speedup builtin
errorsmethod and make it more robust
Datacan no longer be used directly as an array-like object but got mapping-like behavior.- some old deprecated methods were removed
- improved caching speed, reduced tradeoff against memory
- yields were not added correctly in some (especially binned) PDFs and the fit would fail
- add jacobi (many thanks at @HansDembinski for the package)
- reduce the memory footprint on (some) fits, especially repetitive (loops) ones.
Reduces the number of cached compiled functions. The cachesize can be set with
zfit.run.set_cache_size(int)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.
- 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_binnedandto_unbinnedmethods to PDF
- Justin Skorupa for finding the bug in the sum of histograms and the missing automatic conversion of unbinned spaces to binned spaces.
Public release of binned fits and upgrade to Python 3.10 and TensorFlow 2.9.
- improved data handling in constructors
from_pandas(which allows now to have weights as columns, dataframes that are a superset of the obs) andfrom_root(obs can now be spaces and therefore cuts can be direcly applied) - add hashing of unbinned datasets with a
hashintattribute. None if no hash was possible.
- SimpleLoss correctly supports both functions with implicit and explicit parameters, also if they are decorated.
- extended sampling errored for some cases of binned PDFs.
ConstantParametererrored 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
errorsfor 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.
- 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
- Save results by pickling, unpickling a frozen (
FitResult.freeze()) result and usingzfit.param.set_values(params, result)to set the values ofparams.
- the default name of the uncertainty methods
hesseanderrorsdepended 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.
- 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
- 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 usingzfit.param.set_values(params, result)to set the values ofparams.
- params given in ComposedParameters are not sorted anymore. Rely on their name instead.
norm_rangeis now callednormand should be replaced everywhere if possible. This will break in the future.
- remove warning when using
rect_limitsor similar. - gauss integral accepts now also tensor inputs in limits
- parameters at limits is now shown correctly
- add TensorFlow 2.7 support
- fixate nlopt to < 2.7.1
- fixed a longstanding bug in the DoubleCB implementation of the integral.
- remove outdated deprecations
allow
FitResulttofreeze(), 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: - :class:`~zfit.pdf.KDE1DimExact` for the normal density estimation - :class:`~zfit.pdf.KDE1DimGrid` using a binning - :class:`~zfit.pdf.KDE1DimFFT` using a binning and FFT - :class:`~zfit.pdf.KDE1DimISJ` using a binning and an algorithm (ISJ) to solve the optimal bandwidth
For an introduction, see either :ref:`sec-kernel-density-estimation` or the tutorial :ref:`sec-components-model`
add windows in CI
- 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.
- 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.
- add TensorFlow 2.6 support (now 2.5 and 2.6 are supported)
- Marc Steiner for contributing many new KDE methods!
- fix wrong arguments to
minimize - make BaseMinimizer arguments optional
- 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).
- add Python 3.9 support
- upgrade to TensorFlow 2.5
- Scipy minimizers with hessian arguments use now
BFGSas default
- remove Python 3.6 support
- boost-histogram
Update ipyopt requirement < 0.12 to allow numpy compatible with TensorFlow
- hotfix for wrong argument in exponential PDF
- removed requirement ipyopt, can be installed with
pip install zfit[ipyopt]or by manually installingpip install ipyopt
- remove requirement of Ipyopt on MacOS as no wheels are available. This rendered zfit basically non-installable.
- fix loss failed for large datasets
- catch hesse failing for iminuit
Minor small fixes.
- add
lossto callback signature that gives full access to the model - add :meth:`~zfit.loss.UnbinnedNLL.create_new` to losses in order to re-instantiate them with new models and data preserving their current (and future) options and other arguments
Release for fix of minimizers that performed too bad
- remove badly performing Scipy minimizers :class:`~zfit.minimize.ScipyTrustKrylovV1` and :class:`~zfit.minimize.ScipyTrustNCGV1`
- fix auto conversion to complex parameter using constructor
Added many new minimizers from different libraries, all with uncertainty estimation available.
- upgraded to TensorFlow 2.4
- Added many new minimizers. A full list can be found in :ref:`minimize_user_api`.
- :class:`~zfit.minimize.IpyoptV1` that wraps the powerful Ipopt large scale minimization library
- Scipy minimizers now have their own, dedicated wrapper for each instance such as :class:`~zfit.minimize.ScipyLBFGSBV1`, or :class:`~zfit.minimize.ScipySLSQPV1`
- NLopt library wrapper that contains many algorithms for local searches such as :class:`~zfit.minimize.NLoptLBFGSV1`, :class:`~zfit.minimize.NLoptTruncNewtonV1` or :class:`~zfit.minimize.NLoptMMAV1` but also includes more global minimizers such as :class:`~zfit.minimize.NLoptMLSLV1` and :class:`~zfit.minimize.NLoptESCHV1`.
- 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
inittominimize - 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 ofMINOS) minuit_hesseandminuit_minosare now available with all minimizers as well thanks to an great improvement in iminuit.- Added an
approxhesse that returns the approximate hessian (if available, otherwise empty)
- improved
- 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_errorsin speed by caching values and the reliability by making the solution unique. - increased stability for large datasets with a constant subtraction in the NLL
- NLL (and extended) subtracts now by default a constant value. This can be changed with a new
optionsargument. 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, usignset_valueorset_valuesnow raises aValueErrorif the value is outside the limits. Useassignto suppress it.
- strategy to minimizer should now be a class, not an instance anymore.
zfit_errormoved 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_samplernow samples immediately to allow for precompile, a new hook that will allow objects to optimize themselves.
- 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
Update to fix iminuit version
- Fix issue when using a
ComposedParameteras therateargument of aPoissonPDF
- require iminuit < 2 to avoid breaking changes
Upgrade to TensorFlow 2.3 and support for weighted hessian error estimation.
Added a one dimensional Convolution PDF
- upgrad to TensorFlow 2.3
- 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!)
- TensorFlow==2.3 (before 2.2)
tensorflow_probability==0.11- tensorflow-addons # spline interpolation in convolution
- completely new doc design
- 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.loghas been moved toz.math.log(following TF)
- 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_yieldto BasePDF to allow setting the yield in place - Fix possible bug in SumPDF with extended pdfs (automatically)
- upgrade to iminuit>=1.4
- remove cloudpickle hack fix
Johannes for the docs re-design
Kernel density estimation for 1 dimension.
- add correlation method to FitResult
- Gaussian (Truncated) Kernel Density Estimation in one dimension
zfit.pdf.GaussianKDE1DimV1implementation 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.
- change human-readable name of
Gauss,UniformandTruncatedGaussto remove the'_tfp'at the end of the name
- 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
devextra 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_minuitconstructor toFitResultallowing to create it when using directly iminuit - fix possible bias with sampling using accept-reject
- pin down cloudpickle version (upstream bug with pip install) and TF, TFP versions
- Python 3.8 and TF 2.2 support
- easier debugigng with
set_graph_modethat can also be used temporarily with a context manager. False will make everything execute Numpy-like.
- added
get_paramsto loss - fix a bug with the
fixed_paramswhen creating a sampler - improve exponential PDF stability and shift when normalized
- improve accept reject sampling to account for low statistics
- TensorFlow >= 2.2
(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.
Arbitrary limits as well as vectorization (experimental) are now fully supported. The new
Spacehas an additional argument for a function that tests if a vector x is inside.To test if a value is inside a space,
Space.insidecan 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, usingz.unstack_x(lower)be treated likeData. This has also some consequences for the output format ofrect_area: this is now a vector.Due to the ambiguity of the name
limits,areaetc (since they do only reflect the rectangular case) method with leadingrect_*have been added (rect_limits,rect_areaetc.) 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.functioncan now, as withtf.function, be used either as a decorator without arguments or as a decorator with arguments. They are the same as intf.function, except of a few additional ones. zfit.run.set_modeallows 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.pdfwon't trigger a graph build,loss.value()will) withgraph='auto'.- The graph cache can be cleaned manually using
zfit.run.clear_graph_cache()in order to prevent slowness in repeated tasks.
- the
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 invalidFitResult), use theDefaultToyStrategyinzfit.mnimize.Exceptions are now publicly available in
zfit.exceptionAdded nice printout for
FitResultandFitResult.params.get_paramsis now more meaningful, returning by default all independent parameters of the pdf, including yields. Arguments (floating,is_yield) allow for more fine-grained control.
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 functionallyzfit.dimension.add_spacesTo extract limits from multiple limits,
MultiSpaceandSpaceare both iterables, returning the containing spaces respectively itself (for theSpacecase).SumPDF changed in the behavior. Read above in the Major Features and Improvement.
Integrals of extended PDFs are not extended anymore, but
ext_integratenow returns the integral multiplied by the yield.
ComposedParametertakes nowparamsinstead ofdependentsas argument, it acts now as the arguments to thevalue_fn. To stay future compatible, create e.g.def value_fn(p1, pa2)and usingparams = ['param1, param2],value_fnwill then be called asvalue_fn(param1, parma2).value_fnwithout arguments will probably break in the future.FitResult.errorhas been renamed toerrorsto better reflect that multiple errors, the lower and upper are returned.
- 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_valuesmethod that allows to set the values of multiple parameters with one command. Can, as theset_valuemethod be used with a context manager. - wrong size of weights when applying cuts in a dataset
with_coordsdid 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 aFitResultasvaluesargument 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.
- 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.
- refactor
hesse_npwith covariance matrix, make it available to all minimizers
- fix bug in
hesse_np
- Refactoring of the Constraints, dividing into
ProbabilityConstraintthat 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
hesseerror method ('hesse_np') available to all minimizers (not just Minuit).
- Changed default step size to an adaptive scheme, a fraction (1e-4) of the range between the lower and upper limits.
- Add
use_minuit_gradoption 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!
Release to keep up with TensorFlow 2.1
- Fixed the comparison in caching the graph (implementation detail) that leads to an error.
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.
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 aSimpleLossas dependents (assuming the loss function depends on it).ComposedParameterdoes no longer allow to give a Tensor but requires a function that, when evaluated, returns the value. It depends on thedependentsthat are now required.Added numerical differentiation, which allows now to wrap any function with
z.py_function(zfit.z). This can be switched on withzfit.settings.options['numerical_grad'] = TrueAdded 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
zand can be used aszfit.zor imported from it. Added:functiondecorator that can be used to trace a function. Respects dependencies of inputs and automatically caches/invalidates the graph and recreates.py_function, same astf.py_function, but checks and may extends in the futuremathmodule that contains autodiff and numerical differentiation methods, both working with tensors.
- 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_pdfhas troubles with the newest TF version and may not work properly (runs out of memory)
- added numdifftools (for numerical differentiation)
This is a legacy release to add some fixes, next release is TF 2 eager mode only release.
- mostly TF 2.0 compatibility in graph mode, tests against 1.x and 2.x
get_depententsreturns now an OrderedSet- errordef is now a (hidden) attribute and can be changed
- fix bug in polynomials
- added ordered-set
Special release for conda deployment and version fix (TF 2.0 is out)
This is the last release before breaking changes occur
- added ConstantParameter and
zfit.paramnamespace- Available on conda-forge
- 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.
- added native support TFP distributions for analytic sampling
- fix Gaussian (TFP Distribution) Constraint with mixed up order of parameters
from_numpyautomatically converts to default float regardless the original numpy dtype,dtypehas to be used as an explicit argument
- TensorFlow >= 1.14 is required
- Chris Burr for the conda-forge deployment
This is the last release before breaking changes occur
- create
Constraintclass 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)
None
- fixed various small bugs
for the contribution of the Constraints to Matthieu Marinangeli <[email protected]>
Fixed Partial numeric integration
Bugfixes mostly, a few major fixes. Partial numeric integration works now.
- Bugfixes
data_rangecuts 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).
With 0.3.2, bugfixes and three changes in the API/behavior
- tfp distributions wrapping is now different with
dist_kwargsallowing for non-Parameter arguments (like other dists)- sampling allows now for importance sampling (sampler in Model specified differently)
model.samplenow also returns a tensor, being consistent withintegrate
- 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
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
Beta stage and first pip release
- First creation of the package.