RELEASE-NOTES.md
document.
If you want a description of the highlights of this release, check out the release announcement on our new website. Feel free to read it, print it out, and give it to people on the street -- because everybody has to know PyMC 4.0 is officially out 🍾
⚠️ The project was renamed to "PyMC". Now the library is installed as "pip install pymc" and imported likeimport pymc as pm
. See this migration guide for more details.⚠️ Theano-PyMC has been replaced with Aesara, so all external references totheano
andtt
need to be replaced withaesara
andat
, respectively (see 4471).⚠️ Support for JAX and JAX samplers, also allows sampling on GPUs. This benchmark shows speed-ups of up to 11x.⚠️ Random seeding behavior changed (see #5787)!- Sampling results will differ from those of v3 when passing the same
random_seed
as before. They will be consistent across subsequent v4 releases unless mentioned otherwise. - Sampling functions no longer respect user-specified global seeding! Always pass
random_seed
to ensure reproducible behavior. random_seed
now accepts RandomState and Generators besides integers.
- Sampling results will differ from those of v3 when passing the same
⚠️ The GLM submodule was removed, please use Bambi instead.⚠️ PyMC now requires SciPy version>= 1.4.1
(see #4857).
- MvNormalRandomWalk, MvStudentTRandomWalk, GARCH11 and EulerMaruyama distributions (see #4642)
- Nested Mixture distributions (see #5533)
pm.sample_posterior_predictive_w
(see #4807)- Partially observed Multivariate distributions (see #5260)
-
Distributions:
-
Univariate censored distributions are now available via
pm.Censored
. #5169 -
The
CAR
distribution has been added to allow for use of conditional autoregressions which often are used in spatial and network models. -
Added a
logcdf
implementation for the Kumaraswamy distribution (see #4706). -
The
OrderedMultinomial
distribution has been added for use on ordinal data which are aggregated by trial, like multinomial observations, whereasOrderedLogistic
only accepts ordinal data in a disaggregated format, like categorical observations (see #4773). -
The
Polya-Gamma
distribution has been added (see #4531). To make use of this distribution, thepolyagamma>=1.3.1
library must be installed and available in the user's environment. -
pm.DensityDist
can now accept an optionallogcdf
keyword argument to pass in a function to compute the cumulative density function of the distribution (see 5026). -
pm.DensityDist
can now accept an optionalmoment
keyword argument to pass in a function to compute the moment of the distribution (see 5026). -
Added an alternative parametrization,
logit_p
topm.Binomial
andpm.Categorical
distributions (see 5637).
-
-
Model dimensions:
- The dimensionality of model variables can now be parametrized through either of
shape
ordims
(see #4696):- With
shape
the length of dimensions must be given numerically or as scalar AesaraVariables
. Numeric entries inshape
restrict the model variable to the exact length and re-sizing is no longer possible. dims
keeps model variables re-sizeable (for example throughpm.Data
) and leads to well defined coordinates inInferenceData
objects.- An
Ellipsis
(...
) in the last position ofshape
ordims
can be used as short-hand notation for implied dimensions.
- With
- New features for
pm.Data
containers:- With
pm.Data(..., mutable=False)
, or by usingpm.ConstantData()
one can now createTensorConstant
data variables. These can be more performant and compatible in situations where a variable doesn't need to be changed viapm.set_data()
. See #5295. If you do need to change the variable, usepm.Data(..., mutable=True)
, orpm.MutableData()
. - New named dimensions can be introduced to the model via
pm.Data(..., dims=...)
. For mutable data variables (see above) the lengths of these dimensions are symbolic, so they can be re-sized viapm.set_data()
. pm.Data
now passes additional kwargs toaesara.shared
/at.as_tensor
. #5098.
- With
- The length of
dims
in the model is now tracked symbolically throughModel.dim_lengths
(see #4625).
- The dimensionality of model variables can now be parametrized through either of
-
Sampling:
- A small change to the mass matrix tuning methods jitter+adapt_diag (the default) and adapt_diag improves performance early on during tuning for some models. #5004
- New experimental mass matrix tuning method jitter+adapt_diag_grad. #5004
- Support for samplers written in JAX:
- Adding support for numpyro's NUTS sampler via
pymc.sampling_jax.sample_numpyro_nuts()
- Adding support for blackjax's NUTS sampler via
pymc.sampling_jax.sample_blackjax_nuts()
(see #5477) pymc.sampling_jax
samplers supportlog_likelihood
,observed_data
, andsample_stats
in returnedInferenceData
object (see #5189)- Adding support for
pm.Deterministic
inpymc.sampling_jax
(see #5182)
- Adding support for numpyro's NUTS sampler via
-
Miscellaneous:
- The new
pm.find_constrained_prior
function can be used to find optimized prior parameters of a distribution under some constraints (e.g lower and upper bound). See #5231. - Nested models now inherit the parent model's coordinates. #5344
softmax
andlog_softmax
functions added tomath
module (see #5279).- Added the low level
compile_forward_sampling_function
method to compile the aesara function responsible for generating forward samples (see #5759).
- The new
pm.sample(return_inferencedata=True)
is now the default (see #4744).- ArviZ
plots
andstats
wrappers were removed. The functions are now just available by their original names (see #4549 and3.11.2
release notes). pm.sample_posterior_predictive(vars=...)
kwarg was removed in favor ofvar_names
(see #4343).ElemwiseCategorical
step method was removed (see #4701)LKJCholeskyCov
'scompute_corr
keyword argument is now set toTrue
by default (see#5382)- Alternative
sd
keyword argument has been removed from all distributions.sigma
should be used instead (see #5583).
Read on if you're a developer. Or curious. Or both.
pm.Bound
interface no longer accepts a callable class as argument, instead it requires an instantiated distribution (created via the.dist()
API) to be passed as an argument. In addition, Bound no longer returns a class instance but works as a normal PyMC distribution. Finally, it is no longer possible to do predictive random sampling from Bounded variables. Please, consult the new documentation for details on how to use Bounded variables (see 4815).- BART has received various updates (5091, 5177, 5229, 4914) but was removed from the main package in #5566. It is now available from pymc-experimental.
- Removed
AR1
.AR
of order 1 should be used instead. (see 5734). - The
pm.EllipticalSlice
sampler was removed (see #5756). BaseStochasticGradient
was removed (see #5630)pm.Distribution(...).logp(x)
is nowpm.logp(pm.Distribution(...), x)
.pm.Distribution(...).logcdf(x)
is nowpm.logcdf(pm.Distribution(...), x)
.pm.Distribution(...).random(size=x)
is nowpm.draw(pm.Distribution(...), draws=x)
.pm.draw_values(...)
andpm.generate_samples(...)
were removed.pm.fast_sample_posterior_predictive
was removed.pm.sample_prior_predictive
,pm.sample_posterior_predictive
andpm.sample_posterior_predictive_w
now return anInferenceData
object by default, instead of a dictionary (see #5073).pm.sample_prior_predictive
no longer returns transformed variable values by default. Pass them by name invar_names
if you want to obtain these draws (see 4769).pm.sample(trace=...)
no longer acceptsMultiTrace
orlen(.) > 0
traces (see 5019#).- Setting of initial values:
- Setting initial values through
pm.Distribution(testval=...)
is nowpm.Distribution(initval=...)
. Model.update_start_values(...)
was removed. Initial values can be set in theModel.initial_values
dictionary directly.- Test values can no longer be set through
pm.Distribution(testval=...)
and must be assigned manually.
- Setting initial values through
transforms
module is no longer accessible at the root level. It is accessible atpymc.distributions.transforms
(see#5347).logp
,dlogp
, andd2logp
andnojac
variations were removed. UseModel.compile_logp
,compile_dlgop
andcompile_d2logp
withjacobian
keyword instead.pm.DensityDist
no longer accepts thelogp
as its first position argument. It is now an optional keyword argument. If you pass a callable as the first positional argument, aTypeError
will be raised (see 5026).pm.DensityDist
now accepts distribution parameters as positional arguments. Passing them as a dictionary in theobserved
keyword argument is no longer supported and will raise an error (see 5026).- The signature of the
logp
andrandom
functions that can be passed into apm.DensityDist
has been changed (see 5026).
-
Signature and default parameters changed for several distributions:
pm.StudentT
now requires eithersigma
orlam
as kwarg (see #5628)pm.StudentT
now requiresnu
to be specified (no longer defaults to 1) (see #5628)pm.AsymmetricLaplace
positional arguments re-ordered (see #5628)pm.AsymmetricLaplace
now requiresmu
to be specified (no longer defaults to 0) (see #5628)ZeroInflatedPoisson
theta
parameter was renamed tomu
(see #5584).pm.GaussianRandomWalk
initial distribution defaults to zero-centered normal with sigma=100 instead of flat (see#5779)pm.AR
initial distribution defaults to unit normal instead of flat (see#5779)
-
logpt
,logpt_sum
,logp_elemwiset
andnojac
variations were removed. UseModel.logpt(jacobian=True/False, sum=True/False)
instead. -
dlogp_nojact
andd2logp_nojact
were removed. UseModel.dlogpt
andd2logpt
withjacobian=False
instead. -
model.makefn
is now calledModel.compile_fn
, andmodel.fn
was removed. -
Methods starting with
fast_*
, such asModel.fast_logp
, were removed. Same applies toPointFunc
classes -
Model(model=...)
kwarg was removed -
Model(theano_config=...)
kwarg was removed -
Model.size
property was removed (useModel.ndim
instead). -
dims
andcoords
handling: -
Transform.forward
andTransform.backward
signatures changed. -
Changes to the Gaussian Process (GP) submodule (see 5055):
- The
gp.prior(..., shape=...)
kwarg was renamed tosize
. - Multiple methods including
gp.prior
now require explicit kwargs. - For all implementations,
gp.Latent
,gp.Marginal
etc.,cov_func
andmean_func
are required kwargs. - In Windows test conda environment the
mkl
version is fixed to version 2020.4, andmkl-service
is fixed to2.3.0
. This was required forgp.MarginalKron
to function properly. gp.MvStudentT
uses rotated samples fromStudentT
directly now, instead of sampling frompm.Chi2
and then frompm.Normal
.- The "jitter" parameter, or the diagonal noise term added to Gram matrices such that the Cholesky is numerically stable, is now exposed to the user instead of hard-coded. See the function
gp.util.stabilize
. - The
is_observed
arguement forgp.Marginal*
implementations has been deprecated. - In the gp.utils file, the
kmeans_inducing_points
function now passes throughkmeans_kwargs
to scipy's k-means function. - The function
replace_with_values
function has been added togp.utils
. MarginalSparse
has been renamedMarginalApprox
.
- The
-
Removed
MixtureSameFamily
.Mixture
is now capable of handling batched multivariate components (see #5438).
- Switched to the pydata-sphinx-theme
- Updated our documentation tooling to use MyST, MyST-NB, sphinx-design, notfound.extension, sphinx-copybutton and sphinx-remove-toctrees.
- Separated the builds of the example notebooks and of the versioned docs.
- Restructured the documentation to facilitate learning paths
- Updated API docs to document objects at the path users should use to import them
⚠️ Fixed old-time bug in Slice sampler that resulted in biased samples (see #5816).- Removed float128 dtype support (see #4514).
- Logp method of
Uniform
andDiscreteUniform
no longer depends onpymc.distributions.dist_math.bound
for proper evaluation (see #4541). - We now include
cloudpickle
as a required dependency, and no longer depend ondill
(see #4858). - The
incomplete_beta
function inpymc.distributions.dist_math
was replaced byaesara.tensor.betainc
(see 4857). math.log1mexp
andmath.log1mexp_numpy
will expect negative inputs in the future. AFutureWarning
is now raised unlessnegative_input=True
is set (see #4860).- Changed name of
Lognormal
distribution toLogNormal
to harmonize CamelCase usage for distribution names. - Attempt to iterate over MultiTrace will raise NotImplementedError.
- Removed silent normalisation of
p
parameters in Categorical and Multinomial distributions (see #5370).
pm.math.cartesian
can now handle inputs that are themselves >1D (see #4482).- Statistics and plotting functions that were removed in
3.11.0
were brought back, albeit with deprecation warnings if an old naming scheme is used (see #4536). In order to future proof your code, rename these function calls:pm.traceplot
→pm.plot_trace
pm.compareplot
→pm.plot_compare
(here you might need to rename some columns in the input according to thearviz.plot_compare
documentation)pm.autocorrplot
→pm.plot_autocorr
pm.forestplot
→pm.plot_forest
pm.kdeplot
→pm.plot_kde
pm.energyplot
→pm.plot_energy
pm.densityplot
→pm.plot_density
pm.pairplot
→pm.plot_pair
- ⚠ Our memoization mechanism wasn't robust against hash collisions (#4506), sometimes resulting in incorrect values in, for example, posterior predictives. The
pymc.memoize
module was removed and replaced withcachetools
. Thehashable
function andWithMemoization
class were moved topymc.util
(see #4525). pm.make_shared_replacements
now retains broadcasting information which fixes issues with Metropolis samplers (see #4492).
Release manager for 3.11.2: Michael Osthege (@michaelosthege)
- Automatic imputations now also work with
ndarray
data, not justpd.Series
orpd.DataFrame
(see#4439). pymc.sampling_jax.sample_numpyro_nuts
now returns samples from transformed random variables, rather than from the unconstrained representation (see #4427).
- We upgraded to
Theano-PyMC v1.1.2
which includes bugfixes for...- ⚠ a problem with
tt.switch
that affected the behavior of several distributions, including at least the following special cases (see #4448)Bernoulli
when all the observed values were the same (e.g.,[0, 0, 0, 0, 0]
).TruncatedNormal
whensigma
was constant andmu
was being automatically broadcasted to match the shape of observations.
- Warning floods and compiledir locking (see #4444)
- ⚠ a problem with
math.log1mexp_numpy
no longer raises RuntimeWarning when given very small inputs. These were commonly observed during NUTS sampling (see #4428).ScalarSharedVariable
can now be used as an input to other RVs directly (see #4445).pm.sample
andpm.find_MAP
no longer change thestart
argument (see #4458).- Fixed
Dirichlet.logp
method to work with unit batch or event shapes (see #4454). - Bugfix in logp and logcdf methods of
Triangular
distribution (see #4470).
Release manager for 3.11.1: Michael Osthege (@michaelosthege)
This release breaks some APIs w.r.t. 3.10.0
. It also brings some dreadfully awaited fixes, so be sure to go through the (breaking) changes below.
- ⚠ Many plotting and diagnostic functions that were just aliasing ArviZ functions were removed (see 4397). This includes
pm.summary
,pm.traceplot
,pm.ess
and many more! - ⚠ We now depend on
Theano-PyMC
version1.1.0
exactly (see #4405). Major refactorings were done inTheano-PyMC
1.1.0. If you implement customOp
s or interact with Theano in any way yourself, make sure to read the Theano-PyMC 1.1.0 release notes. - ⚠ Python 3.6 support was dropped (by no longer testing) and Python 3.9 was added (see #4332).
- ⚠ Changed shape behavior: No longer collapse length 1 vector shape into scalars. (see #4206 and #4214)
- Applies to random variables and also the
.random(size=...)
kwarg! - To create scalar variables you must now use
shape=None
orshape=()
. shape=(1,)
andshape=1
now become vectors. Previously they were collapsed into scalars- 0-length dimensions are now ruled illegal for random variables and raise a
ValueError
.
- Applies to random variables and also the
- In
sample_prior_predictive
thevars
kwarg was removed in favor ofvar_names
(see #4327). - Removed
theanof.set_theano_config
because it illegally changed Theano's internal state (see #4329).
- Option to set
check_bounds=False
when instantiatingpymc.Model()
. This turns off bounds checks that ensure that input parameters of distributions are valid. For correctly specified models, this is unnecessary as all parameters get automatically transformed so that all values are valid. Turning this off should lead to faster sampling (see #4377). OrderedProbit
distribution added (see #4232).plot_posterior_predictive_glm
now works witharviz.InferenceData
as well (see #4234)- Add
logcdf
method to all univariate discrete distributions (see #4387). - Add
random
method toMvGaussianRandomWalk
(see #4388) AsymmetricLaplace
distribution added (see #4392).DirichletMultinomial
distribution added (see #4373).- Added a new
predict
method toBART
to compute out of sample predictions (see #4310).
- Fixed bug whereby partial traces returns after keyboard interrupt during parallel sampling had fewer draws than would've been available #4318
- Make
sample_shape
same across all contexts indraw_values
(see #4305). - The notebook gallery has been moved to https://github.com/pymc-devs/pymc-examples (see #4348).
math.logsumexp
now matchesscipy.special.logsumexp
when arrays contain infinite values (see #4360).- Fixed mathematical formulation in
MvStudentT
random method. (see #4359) - Fix issue in
logp
method ofHyperGeometric
. It now returns-inf
for invalid parameters (see 4367) - Fixed
MatrixNormal
random method to work with parameters as random variables. (see #4368) - Update the
logcdf
method of several continuous distributions to return -inf for invalid parameters and values, and raise an informative error when multiple values cannot be evaluated in a single call. (see 4393 and #4421) - Improve numerical stability in
logp
andlogcdf
methods ofExGaussian
(see #4407) - Issue UserWarning when doing prior or posterior predictive sampling with models containing Potential factors (see #4419)
- Dirichlet distribution's
random
method is now optimized and gives outputs in correct shape (see #4416) - Attempting to sample a named model with SMC will now raise a
NotImplementedError
. (see #4365)
Release manager for 3.11.0: Eelke Spaak (@Spaak)
This is a major release with many exciting new features. The biggest change is that we now rely on our own fork of Theano-PyMC. This is in line with our big announcement about our commitment to PyMC3 and Theano.
When upgrading, make sure that Theano-PyMC
and not Theano
are installed (the imports remain unchanged, however). If not, you can uninstall Theano
:
conda remove theano
And to install:
conda install -c conda-forge theano-pymc
Or, if you are using pip (not recommended):
pip uninstall theano
And to install:
pip install theano-pymc
This new version of Theano-PyMC
comes with an experimental JAX backend which, when combined with the new and experimental JAX samplers in PyMC3, can greatly speed up sampling in your model. As this is still very new, please do not use it in production yet but do test it out and let us know if anything breaks and what results you are seeing, especially speed-wise.
- New experimental JAX samplers in
pymc.sample_jax
(see notebook and #4247). Requires JAX and either TFP or numpyro. - Add MLDA, a new stepper for multilevel sampling. MLDA can be used when a hierarchy of approximate posteriors of varying accuracy is available, offering improved sampling efficiency especially in high-dimensional problems and/or where gradients are not available (see #3926)
- Add Bayesian Additive Regression Trees (BARTs) #4183)
- Added
pymc.gp.cov.Circular
kernel for Gaussian Processes on circular domains, e.g. the unit circle (see #4082). - Added a new
MixtureSameFamily
distribution to handle mixtures of arbitrary dimensions in vectorized form for improved speed (see #4185). sample_posterior_predictive_w
can now feed onxarray.Dataset
- e.g. fromInferenceData.posterior
. (see #4042)- Change SMC metropolis kernel to independent metropolis kernel #4115)
- Add alternative parametrization to NegativeBinomial distribution in terms of n and p (see #4126)
- Added semantically meaningful
str
representations to PyMC3 objects for console, notebook, and GraphViz use (see #4076, #4065, #4159, #4217, #4243, and #4260). - Add Discrete HyperGeometric Distribution (see #4249)
- Switch the dependency of Theano to our own fork, Theano-PyMC.
- Removed non-NDArray (Text, SQLite, HDF5) backends and associated tests.
- Use dill to serialize user defined logp functions in
DensityDist
. The previous serialization code fails if it is used in notebooks on Windows and Mac.dill
is now a required dependency. (see #3844). - Fixed numerical instability in ExGaussian's logp by preventing
logpow
from returning-inf
(see #4050). - Numerically improved stickbreaking transformation - e.g. for the
Dirichlet
distribution. #4129 - Enabled the
Multinomial
distribution to handle batch sizes that have more than 2 dimensions. #4169 - Test model logp before starting any MCMC chains (see #4211)
- Fix bug in
model.check_test_point
that caused thetest_point
argument to be ignored. (see PR #4211) - Refactored MvNormal.random method with better handling of sample, batch and event shapes. #4207
- The
InverseGamma
distribution now implements alogcdf
. #3944 - Make starting jitter methods for nuts sampling more robust by resampling values that lead to non-finite probabilities. A new optional argument
jitter-max-retries
can be passed topm.sample()
andpm.init_nuts()
to control the maximum number of retries per chain. 4298
- Added a new notebook demonstrating how to incorporate sampling from a conjugate Dirichlet-multinomial posterior density in conjunction with other step methods (see #4199).
- Mentioned the way to do any random walk with
theano.tensor.cumsum()
inGaussianRandomWalk
docstrings (see #4048).
Release manager for 3.10.0: Eelke Spaak (@Spaak)
- Introduce optional arguments to
pm.sample
:mp_ctx
to control how the processes for parallel sampling are started, andpickle_backend
to specify which library is used to pickle models in parallel sampling when the multiprocessing context is not of typefork
(see #3991). - Add sampler stats
process_time_diff
,perf_counter_diff
andperf_counter_start
, that record wall and CPU times for each NUTS and HMC sample (see #3986). - Extend
keep_size
argument handling forsample_posterior_predictive
andfast_sample_posterior_predictive
, to work on ArviZInferenceData
and xarrayDataset
input values (see PR #4006 and issue #4004). - SMC-ABC: add the Wasserstein and energy distance functions. Refactor API, the distance, sum_stats and epsilon arguments are now passed
pm.Simulator
instead ofpm.sample_smc
. Add random method topm.Simulator
. Add option to save the simulated data. Improved LaTeX representation #3996. - SMC-ABC: Allow use of potentials by adding them to the prior term. #4016.
- Fix an error on Windows and Mac where error message from unpickling models did not show up in the notebook, or where sampling froze when a worker process crashed (see #3991).
- Require Theano >= 1.0.5 (see #4032).
- Notebook on multilevel modeling has been rewritten to showcase ArviZ and xarray usage for inference result analysis (see #3963).
NB: The docs/*
folder is still removed from the tarball due to an upload size limit on PyPi.
Release manager for 3.9.3: Kyle Beauchamp (@kyleabeauchamp)
- Warning added in GP module when
input_dim
is lower than the number of columns inX
to compute the covariance function (see #3974). - Pass the
tune
argument fromsample
when usingadvi+adapt_diag_grad
(see issue #3965, fixed by #3979). - Add simple test case for new coords and dims feature in
pm.Model
(see #3977). - Require ArviZ >= 0.9.0 (see #3977).
- Fixed issue #3962 by making a change in the
_random()
method ofGaussianRandomWalk
class (see PR #3985). Further testing revealed a new issue which is being tracked by #4010.
NB: The docs/*
folder is still removed from the tarball due to an upload size limit on PyPi.
Release manager for 3.9.2: Alex Andorra (@AlexAndorra)
The v3.9.0
upload to PyPI didn't include a tarball, which is fixed in this release.
Though we had to temporarily remove the docs/*
folder from the tarball due to a size limit.
Release manager for 3.9.1: Michael Osthege (@michaelosthege)
- Use fastprogress instead of tqdm #3693.
DEMetropolis
can now tune bothlambda
andscaling
parameters, but by default neither of them are tuned. See #3743 for more info.DEMetropolisZ
, an improved variant ofDEMetropolis
brings better parallelization and higher efficiency with fewer chains with a slower initial convergence. This implementation is experimental. See #3784 for more info.- Notebooks that give insight into
DEMetropolis
,DEMetropolisZ
and theDifferentialEquation
interface are now located in the Tutorials/Deep Dive section. - Add
fast_sample_posterior_predictive
, a vectorized alternative tosample_posterior_predictive
. This alternative is substantially faster for large models. - GP covariance functions can now be exponentiated by a scalar. See PR #3852
sample_posterior_predictive
can now feed onxarray.Dataset
- e.g. fromInferenceData.posterior
. (see #3846)SamplerReport
(MultiTrace.report
) now has propertiesn_tune
,n_draws
,t_sampling
for increased convenience (see #3827)pm.sample(..., return_inferencedata=True)
can now directly return the trace asarviz.InferenceData
(see #3911)pm.sample
now has support for adapting dense mass matrix usingQuadPotentialFullAdapt
(see #3596, #3705, #3858, and #3893). Useinit="adapt_full"
orinit="jitter+adapt_full"
to use.Moyal
distribution added (see #3870).pm.LKJCholeskyCov
now automatically computes and returns the unpacked Cholesky decomposition, the correlations and the standard deviations of the covariance matrix (see #3881).pm.Data
container can now be used for index variables, i.e with integer data and not only floats (issue #3813, fixed by #3925).pm.Data
container can now be used as input for other random variables (issue #3842, fixed by #3925).- Allow users to specify coordinates and dimension names instead of numerical shapes when specifying a model. This makes interoperability with ArviZ easier. (see #3551)
- Plots and Stats API sections now link to ArviZ documentation #3927
- Add
SamplerReport
with propertiesn_draws
,t_sampling
andn_tune
to SMC.n_tune
is always 0 #3931. - SMC-ABC: add option to define summary statistics, allow to sample from more complex models, remove redundant distances #3940
- Tuning results no longer leak into sequentially sampled
Metropolis
chains (see #3733 and #3796). - We'll deprecate the
Text
andSQLite
backends and thesave_trace
/load_trace
functions, since this is now done with ArviZ. (see #3902) - ArviZ
v0.8.3
is now the minimum required version - In named models,
pm.Data
objects now get model-relative names (see #3843). -
pm.sample
now takes 1000 draws and 1000 tuning samples by default, instead of 500 previously (see #3855). - Moved argument division out of
NegativeBinomial
random
method. Fixes #3864 in the style of #3509. - The Dirichlet distribution now raises a ValueError when it's initialized with <= 0 values (see #3853).
- Dtype bugfix in
MvNormal
andMvStudentT
(see 3836). - End of sampling report now uses
arviz.InferenceData
internally and avoids storing pointwise log likelihood (see #3883). - The multiprocessing start method on MacOS is now set to "forkserver", to avoid crashes (see issue #3849, solved by #3919).
- The AR1 logp now uses the precision of the whole AR1 process instead of just the innovation precision (see issue #3892, fixed by #3899).
- Forced the
Beta
distribution'srandom
method to generate samples that are in the open interval$(0, 1)$ , i.e. no value can be equal to zero or equal to one (issue #3898 fixed by #3924). - Fixed an issue that happened on Windows, that was introduced by the clipped beta distribution rvs function (#3924). Windows does not support the
float128
dtype, but we had assumed that it had to be available. The solution was to only supportfloat128
on Linux and Darwin systems (see issue #3929 fixed by #3930).
- Remove
sample_ppc
andsample_ppc_w
that were deprecated in 3.6. - Deprecated
sd
has been replaced bysigma
(already in version 3.7) in continuous, mixed and timeseries distributions and now raisesDeprecationWarning
whensd
is used. (see #3837 and #3688). - We'll deprecate the
Text
andSQLite
backends and thesave_trace
/load_trace
functions, since this is now done with ArviZ. (see #3902) - Dropped some deprecated kwargs and functions (see #3906)
- Dropped the outdated 'nuts' initialization method for
pm.sample
(see #3863).
Release manager for 3.9.0: Michael Osthege (@michaelosthege)
- Implemented robust u turn check in NUTS (similar to stan-dev/stan#2800). See PR [#3605]
- Add capabilities to do inference on parameters in a differential equation with
DifferentialEquation
. See #3590 and #3634. - Distinguish between
Data
andDeterministic
variables when graphing models with graphviz. PR #3491. - Sequential Monte Carlo - Approximate Bayesian Computation step method is now available. The implementation is in an experimental stage and will be further improved.
- Added
Matern12
covariance function for Gaussian processes. This is the Matern kernel with nu=1/2. - Progressbar reports number of divergences in real time, when available #3547.
- Sampling from variational approximation now allows for alternative trace backends [#3550].
- Infix
@
operator now works with random variables and deterministics #3619. - ArviZ is now a requirement, and handles plotting, diagnostics, and statistical checks.
- Can use GaussianRandomWalk in sample_prior_predictive and sample_prior_predictive #3682
- Now 11 years of S&P returns in data set#3682
- Moved math operations out of
Rice
,TruncatedNormal
,Triangular
andZeroInflatedNegativeBinomial
random
methods. Math operations on values returned bydraw_values
might not broadcast well, and all thesize
aware broadcasting is left togenerate_samples
. Fixes #3481 and #3508 - Parallelization of population steppers (
DEMetropolis
) is now set via thecores
argument. (#3559) - Fixed a bug in
Categorical.logp
. In the case of multidimensionalp
's, the indexing was done wrong leading to incorrectly shaped tensors that consumedO(n**2)
memory instead ofO(n)
. This fixes issue #3535 - Fixed a defect in
OrderedLogistic.__init__
that unnecessarily increased the dimensionality of the underlyingp
. Related to issue issue #3535 but was not the true cause of it. - SMC: stabilize covariance matrix 3573
- SMC: is no longer a step method of
pm.sample
now it should be called usingpm.sample_smc
3579 - SMC: improve computation of the proposal scaling factor 3594 and 3625
- SMC: reduce number of logp evaluations 3600
- SMC: remove
scaling
andtune_scaling
arguments as is a better idea to always allow SMC to automatically compute the scaling factor 3625 - Now uses
multiprocessong
rather thanpsutil
to count CPUs, which results in reliable core counts on Chromebooks. sample_posterior_predictive
now preallocates the memory required for its output to improve memory usage. Addresses problems raised in this discourse thread.- Fixed a bug in
Categorical.logp
. In the case of multidimensionalp
's, the indexing was done wrong leading to incorrectly shaped tensors that consumedO(n**2)
memory instead ofO(n)
. This fixes issue #3535 - Fixed a defect in
OrderedLogistic.__init__
that unnecessarily increased the dimensionality of the underlyingp
. Related to issue issue #3535 but was not the true cause of it. - Wrapped
DensityDist.rand
withgenerate_samples
to make it aware of the distribution's shape. Added control flow attributes to still be able to behave as in earlier versions, and to control how to interpret thesize
parameter in therandom
callable signature. Fixes 3553 - Added
theano.gof.graph.Constant
to type checks done in_draw_value
(fixes issue 3595) HalfNormal
did not used to work properly indraw_values
,sample_prior_predictive
, orsample_posterior_predictive
(fixes issue 3686)- Random variable transforms were inadvertently left out of the API documentation. Added them. (See PR 3690).
- Refactored
pymc.model.get_named_nodes_and_relations
to use the ancestors and descendents, in a way that is consistent withtheano
's naming convention. - Changed the way in which
pymc.model.get_named_nodes_and_relations
computes nodes without ancestors to make it robust to changes in var_name orderings (issue #3643)
- Add data container class (
Data
) that wraps the theano SharedVariable class and let the model be aware of its inputs and outputs. - Add function
set_data
to update variables defined asData
. Mixture
now supports mixtures of multidimensional probability distributions, not just lists of 1D distributions.GLM.from_formula
andLinearComponent.from_formula
can extract variables from the calling scope. Customizable via the neweval_env
argument. Fixing #3382.- Added the
distributions.shape_utils
module with functions used to help broadcast samples drawn from distributions using thesize
keyword argument. - Used
numpy.vectorize
indistributions.distribution._compile_theano_function
. This enablessample_prior_predictive
andsample_posterior_predictive
to ask for tuples of samples instead of just integers. This fixes issue #3422.
- All occurrences of
sd
as a parameter name have been renamed tosigma
.sd
will continue to function for backwards compatibility. HamiltonianMC
was ignoring certain arguments liketarget_accept
, and not using the custom step size jitter function with expectation 1.- Made
BrokenPipeError
for parallel sampling more verbose on Windows. - Added the
broadcast_distribution_samples
function that helps broadcasting arrays of drawn samples, taking into account the requestedsize
and the inferred distribution shape. This sometimes is needed by distributions that call severalrvs
separately within theirrandom
method, such as theZeroInflatedPoisson
(fixes issue #3310). - The
Wald
,Kumaraswamy
,LogNormal
,Pareto
,Cauchy
,HalfCauchy
,Weibull
andExGaussian
distributionsrandom
method used a hidden_random
function that was written with scalars in mind. This could potentially lead to artificial correlations between random draws. Added shape guards and broadcasting of the distribution samples to prevent this (Similar to issue #3310). - Added a fix to allow the imputation of single missing values of observed data, which previously would fail (fixes issue #3122).
- The
draw_values
function was too permissive with what could be grabbed from insidepoint
, which lead to an error when sampling posterior predictives of variables that depended on shared variables that had changed their shape afterpm.sample()
had been called (fix issue #3346). draw_values
now adds the theano graph descendants ofTensorConstant
orSharedVariables
to the named relationship nodes stack, only if these descendants areObservedRV
orMultiObservedRV
instances (fixes issue #3354).- Fixed bug in broadcast_distrution_samples, which did not handle correctly cases in which some samples did not have the size tuple prepended.
- Changed
MvNormal.random
's usage oftensordot
for Cholesky encoded covariances. This lead to wrong axis broadcasting and seemed to be the cause for issue #3343. - Fixed defect in
Mixture.random
when multidimensional mixtures were involved. The mixture component was not preserved across all the elements of the dimensions of the mixture. This meant that the correlations across elements within a given draw of the mixture were partly broken. - Restructured
Mixture.random
to allow better use of vectorized calls tocomp_dists.random
. - Added tests for mixtures of multidimensional distributions to the test suite.
- Fixed incorrect usage of
broadcast_distribution_samples
inDiscreteWeibull
. Mixture
's default dtype is now determined bytheano.config.floatX
.dist_math.random_choice
now handles nd-arrays of category probabilities, and also handles sizes that are notNone
. Also removed unusedk
kwarg fromdist_math.random_choice
.- Changed
Categorical.mode
to preserve all the dimensions ofp
except the last one, which encodes each category's probability. - Changed initialization of
Categorical.p
.p
is now normalized to sum to1
insidelogp
andrandom
, but not during initialization. This could hide negative values supplied top
as mentioned in #2082. Categorical
now accepts elements ofp
equal to0
.logp
will return-inf
if there arevalues
that index to the zero probability categories.- Add
sigma
,tau
, andsd
to signature ofNormalMixture
. - Set default lower and upper values of -inf and inf for pm.distributions.continuous.TruncatedNormal. This avoids errors caused by their previous values of None (fixes issue #3248).
- Converted all calls to
pm.distributions.bound._ContinuousBounded
andpm.distributions.bound._DiscreteBounded
to use only and all positional arguments (fixes issue #3399). - Restructured
distributions.distribution.generate_samples
to use theshape_utils
module. This solves issues #3421 and #3147 by using thesize
aware broadcating functions inshape_utils
. - Fixed the
Multinomial.random
andMultinomial.random_
methods to make them compatible with the newgenerate_samples
function. In the process, a bug of theMultinomial.random_
shape handling was discovered and fixed. - Fixed a defect found in
Bound.random
where thepoint
dictionary was passed togenerate_samples
as anarg
instead of innot_broadcast_kwargs
. - Fixed a defect found in
Bound.random_
wheretotal_size
could end up as afloat64
instead of being an integer if givensize=tuple()
. - Fixed an issue in
model_graph
that caused construction of the graph of the model for rendering to hang: replaced a search over the powerset of the nodes with a breadth-first search over the nodes. Fix for #3458. - Removed variable annotations from
model_graph
but left type hints (Fix for #3465). This means that we supportpython>=3.5.4
. - Default
target_accept
forHamiltonianMC
is now 0.65, as suggested in Beskos et. al. 2010 and Neal 2001. - Fixed bug in
draw_values
that lead to intermittent errors in python3.5. This happened with some deterministic nodes that were drawn but not added togivens
.
nuts_kwargs
andstep_kwargs
have been deprecated in favor of using the standardkwargs
to pass optional step method arguments.SGFS
andCSG
have been removed (Fix for #3353). They have been moved to pymc-experimental.- References to
live_plot
and corresponding notebooks have been removed. - Function
approx_hessian
was removed, due tonumdifftools
becoming incompatible with currentscipy
. The function was already optional, only available to a user who installednumdifftools
separately, and not hit on any common codepaths. #3485. - Deprecated
vars
parameter ofsample_posterior_predictive
in favor ofvarnames
. - References to
live_plot
and corresponding notebooks have been removed. - Deprecated
vars
parameters ofsample_posterior_predictive
andsample_prior_predictive
in favor ofvar_names
. At least for the latter, this is more accurate, since thevars
parameter actually took names.
45 Luciano Paz
38 Thomas Wiecki
23 Colin Carroll
19 Junpeng Lao
15 Chris Fonnesbeck
13 Juan Martín Loyola
13 Ravin Kumar
8 Robert P. Goldman
5 Tim Blazina
4 chang111
4 adamboche
3 Eric Ma
3 Osvaldo Martin
3 Sanmitra Ghosh
3 Saurav Shekhar
3 chartl
3 fredcallaway
3 Demetri
2 Daisuke Kondo
2 David Brochart
2 George Ho
2 Vaibhav Sinha
1 rpgoldman
1 Adel Tomilova
1 Adriaan van der Graaf
1 Bas Nijholt
1 Benjamin Wild
1 Brigitta Sipocz
1 Daniel Emaasit
1 Hari
1 Jeroen
1 Joseph Willard
1 Juan Martin Loyola
1 Katrin Leinweber
1 Lisa Martin
1 M. Domenzain
1 Matt Pitkin
1 Peadar Coyle
1 Rupal Sharma
1 Tom Gilliss
1 changjiangeng
1 michaelosthege
1 monsta
1 579397
This will be the last release to support Python 2.
- Track the model log-likelihood as a sampler stat for NUTS and HMC samplers
(accessible as
trace.get_sampler_stats('model_logp')
) (#3134) - Add Incomplete Beta function
incomplete_beta(a, b, value)
- Add log CDF functions to continuous distributions:
Beta
,Cauchy
,ExGaussian
,Exponential
,Flat
,Gumbel
,HalfCauchy
,HalfFlat
,HalfNormal
,Laplace
,Logistic
,LogNormal
,Normal
,Pareto
,StudentT
,Triangular
,Uniform
,Wald
,Weibull
. - Behavior of
sample_posterior_predictive
is now to produce posterior predictive samples, in order, from all values of thetrace
. Previously, by default it would produce 1 chain worth of samples, using a random selection from thetrace
(#3212) - Show diagnostics for initial energy errors in HMC and NUTS.
- PR #3273 has added the
distributions.distribution._DrawValuesContext
context manager. This is used to store the values already drawn in nestedrandom
anddraw_values
calls, enablingdraw_values
to draw samples from the joint probability distribution of RVs and not the marginals. Custom distributions that must calldraw_values
several times in theirrandom
method, or that invoke many calls to other distribution'srandom
methods (e.g. mixtures) must do all of these calls under the same_DrawValuesContext
context manager instance. If they do not, the conditional relations between the distribution's parameters could be broken, andrandom
could return values drawn from an incorrect distribution. Rice
distribution is now defined with either the noncentrality parameter or the shape parameter (#3287).
- Big rewrite of documentation (#3275)
- Fixed Triangular distribution
c
attribute handling inrandom
and updated sample codes for consistency (#3225) - Refactor SMC and properly compute marginal likelihood (#3124)
- Removed use of deprecated
ymin
keyword in matplotlib'sAxes.set_ylim
(#3279) - Fix for #3210. Now
distribution.draw_values(params)
, will draw theparams
values from their joint probability distribution and not from combinations of their marginals (Refer to PR #3273). - Removed dependence on pandas-datareader for retrieving Yahoo Finance data in examples (#3262)
- Rewrote
Multinomial._random
method to better handle shape broadcasting (#3271) - Fixed
Rice
distribution, which inconsistently mixed two parametrizations (#3286). Rice
distribution now accepts multiple parameters and observations and is usable with NUTS (#3289).sample_posterior_predictive
no longer callsdraw_values
to initialize the shape of the ppc trace. This called could lead toValueError
's when sampling the ppc from a model withFlat
orHalfFlat
prior distributions (Fix issue #3294).- Added explicit conversion to
floatX
andint32
for the continuous and discrete probability distribution parameters (addresses issue #3223).
- Renamed
sample_ppc()
andsample_ppc_w()
tosample_posterior_predictive()
andsample_posterior_predictive_w()
, respectively.
- Add documentation section on survival analysis and censored data models
- Add
check_test_point
method topm.Model
- Add
Ordered
Transformation andOrderedLogistic
distribution - Add
Chain
transformation - Improve error message
Mass matrix contains zeros on the diagonal. Some derivatives might always be zero
during tuning ofpm.sample
- Improve error message
NaN occurred in optimization.
during ADVI - Save and load traces without
pickle
usingpm.save_trace
andpm.load_trace
- Add
Kumaraswamy
distribution - Add
TruncatedNormal
distribution - Rewrite parallel sampling of multiple chains on py3. This resolves long standing issues when transferring large traces to the main process, avoids pickling issues on UNIX, and allows us to show a progress bar for all chains. If parallel sampling is interrupted, we now return partial results.
- Add
sample_prior_predictive
which allows for efficient sampling from the unconditioned model. - SMC: remove experimental warning, allow sampling using
sample
, reduce autocorrelation from final trace. - Add
model_to_graphviz
(which uses the optional dependencygraphviz
) to plot a directed graph of a PyMC3 model using plate notation. - Add beta-ELBO variational inference as in beta-VAE model (Christopher P. Burgess et al. NIPS, 2017)
- Add
__dir__
toSingleGroupApproximation
to improve autocompletion in interactive environments
- Fixed grammar in divergence warning, previously
There were 1 divergences ...
could be raised. - Fixed
KeyError
raised when only subset of variables are specified to be recorded in the trace. - Removed unused
repeat=None
arguments from allrandom()
methods in distributions. - Deprecated the
sigma
argument inMarginalSparse.marginal_likelihood
in favor ofnoise
- Fixed unexpected behavior in
random
. Now therandom
functionality is more robust and will work better forsample_prior
when that is implemented. - Fixed
scale_cost_to_minibatch
behaviour, previously this was not working and alwaysFalse
- Add
logit_p
keyword topm.Bernoulli
, so that users can specify the logit of the success probability. This is faster and more stable than usingp=tt.nnet.sigmoid(logit_p)
. - Add
random
keyword topm.DensityDist
thus enabling users to pass custom random method which in turn makes sampling from aDensityDist
possible. - Effective sample size computation is updated. The estimation uses Geyer's initial positive sequence, which no longer truncates the autocorrelation series inaccurately.
pm.diagnostics.effective_n
now can reports N_eff>N. - Added
KroneckerNormal
distribution and a correspondingMarginalKron
Gaussian Process implementation for efficient inference, along with lower-level functions such ascartesian
andkronecker
products. - Added
Coregion
covariance function. - Add new 'pairplot' function, for plotting scatter or hexbin matrices of sampled parameters. Optionally it can plot divergences.
- Plots of discrete distributions in the docstrings
- Add logitnormal distribution
- Densityplot: add support for discrete variables
- Fix the Binomial likelihood in
.glm.families.Binomial
, with the flexibility of specifying then
. - Add
offset
kwarg to.glm
. - Changed the
compare
function to accept a dictionary of model-trace pairs instead of two separate lists of models and traces. - add test and support for creating multivariate mixture and mixture of mixtures
distribution.draw_values
, now is also able to draw values from conditionally dependent RVs, such as autotransformed RVs (Refer to PR #2902).
VonMises
does not overflow for large values of kappa. i0 and i1 have been removed and we now use log_i0 to compute the logp.- The bandwidth for KDE plots is computed using a modified version of Scott's rule. The new version uses entropy instead of standard deviation. This works better for multimodal distributions. Functions using KDE plots has a new argument
bw
controlling the bandwidth. - fix PyMC3 variable is not replaced if provided in more_replacements (#2890)
- Fix for issue #2900. For many situations, named node-inputs do not have a
random
method, while some intermediate node may have it. This meant that if the named node-input at the leaf of the graph did not have a fixed value,theano
would try to compile it and fail to find inputs, raising atheano.gof.fg.MissingInputError
. This was fixed by going through the theano variable's owner inputs graph, trying to get intermediate named-nodes values if the leafs had failed. - In
distribution.draw_values
, some named nodes could betheano.tensor.TensorConstant
s ortheano.tensor.sharedvar.SharedVariable
s. Nevertheless, indistribution._draw_value
, these would be passed todistribution._compile_theano_function
as if they weretheano.tensor.TensorVariable
s. This could lead to the following exceptionsTypeError: ('Constants not allowed in param list', ...)
orTypeError: Cannot use a shared variable (...)
. The fix was to not addtheano.tensor.TensorConstant
ortheano.tensor.sharedvar.SharedVariable
named nodes into thegivens
dict that could be used indistribution._compile_theano_function
. - Exponential support changed to include zero values.
- DIC and BPIC calculations have been removed
- df_summary have been removed, use summary instead
njobs
andnchains
kwarg are deprecated in favor ofcores
andchains
forsample
lag
kwarg inpm.stats.autocorr
andpm.stats.autocov
is deprecated.
- Improve NUTS initialization
advi+adapt_diag_grad
and addjitter+adapt_diag_grad
(#2643) - Added
MatrixNormal
class for representing vectors of multivariate normal variables - Implemented
HalfStudentT
distribution - New benchmark suite added (see http://pandas.pydata.org/speed/pymc/)
- Generalized random seed types
- Update loo, new improved algorithm (#2730)
- New CSG (Constant Stochastic Gradient) approximate posterior sampling algorithm (#2544)
- Michael Osthege added support for population-samplers and implemented differential evolution metropolis (
DEMetropolis
). For models with correlated dimensions that can not use gradient-based samplers, theDEMetropolis
sampler can give higher effective sampling rates. (also see PR#2735) - Forestplot supports multiple traces (#2736)
- Add new plot, densityplot (#2741)
- DIC and BPIC calculations have been deprecated
- Refactor HMC and implemented new warning system (#2677, #2808)
- Fixed
compareplot
to useloo
output. - Improved
posteriorplot
to scale fonts sample_ppc_w
now broadcastsdf_summary
function renamed tosummary
- Add test for
model.logp_array
andmodel.bijection
(#2724) - Fixed
sample_ppc
andsample_ppc_w
to iterate all chains(#2633, #2748) - Add Bayesian R2 score (for GLMs)
stats.r2_score
(#2696) and test (#2729). - SMC works with transformed variables (#2755)
- Speedup OPVI (#2759)
- Multiple minor fixes and improvements in the docs (#2775, #2786, #2787, #2789, #2790, #2794, #2799, #2809)
- Old (
minibatch-
)advi
is removed (#2781)
This version includes two major contributions from our Google Summer of Code 2017 students:
- Maxim Kochurov extended and refactored the variational inference module. This primarily adds two important classes, representing operator variational inference (
OPVI
) objects andApproximation
objects. These make it easier to extend existingvariational
classes, and to derive inference fromvariational
optimizations, respectively. Thevariational
module now also includes normalizing flows (NFVI
). - Bill Engels added an extensive new Gaussian processes (
gp
) module. Standard GPs can be specified using eitherLatent
orMarginal
classes, depending on the nature of the underlying function. A Student-T processTP
has been added. In order to accommodate larger datasets, approximate marginal Gaussian processes (MarginalSparse
) have been added.
Documentation has been improved as the result of the project's monthly "docathons".
An experimental stochastic gradient Fisher scoring (SGFS
) sampling step method has been added.
The API for find_MAP
was enhanced.
SMC now estimates the marginal likelihood.
Added Logistic
and HalfFlat
distributions to set of continuous distributions.
Bayesian fraction of missing information (bfmi
) function added to stats
.
Enhancements to compareplot
added.
QuadPotential adaptation has been implemented.
Script added to build and deploy documentation.
MAP estimates now available for transformed and non-transformed variables.
The Constant
variable class has been deprecated, and will be removed in 3.3.
DIC and BPIC calculations have been sped up.
Arrays are now accepted as arguments for the Bound
class.
random
method was added to the Wishart
and LKJCorr
distributions.
Progress bars have been added to LOO and WAIC calculations.
All example notebooks updated to reflect changes in API since 3.1.
Parts of the test suite have been refactored.
Fixed sampler stats error in NUTS for non-RAM backends
Matplotlib is no longer a hard dependency, making it easier to use in settings where installing Matplotlib is problematic. PyMC3 will only complain if plotting is attempted.
Several bugs in the Gaussian process covariance were fixed.
All chains are now used to calculate WAIC and LOO.
AR(1) log-likelihood function has been fixed.
Slice sampler fixed to sample from 1D conditionals.
Several docstring fixes.
The following people contributed to this release (ordered by number of commits):
Maxim Kochurov [email protected] Bill Engels [email protected] Chris Fonnesbeck [email protected] Junpeng Lao [email protected] Adrian Seyboldt [email protected] AustinRochford [email protected] Osvaldo Martin [email protected] Colin Carroll [email protected] Hannes Vasyura-Bathke [email protected] Thomas Wiecki [email protected] michaelosthege [email protected] Marco De Nadai [email protected] Kyle Beauchamp [email protected] Massimo [email protected] ctm22396 [email protected] Max Horn [email protected] Hennadii Madan [email protected] Hassan Naseri [email protected] Peadar Coyle [email protected] Saurav R. Tuladhar [email protected] Shashank Shekhar [email protected] Eric Ma [email protected] Ed Herbst [email protected] tsdlovell [email protected] zaxtax [email protected] Dan Nichol [email protected] Benjamin Yetton [email protected] jackhansom [email protected] Jack Tsai [email protected] Andrés Asensio Ramos [email protected]
-
New user forum at http://discourse.pymc.io
-
Much improved variational inference support:
-
Add Stein-Variational Gradient Descent as well as Amortized SVGD (experimental).
-
Added various optimizers including ADAM.
-
Stopping criterion implemented via callbacks.
-
sample() defaults changed: tuning is enabled for the first 500 samples which are then discarded from the trace as burn-in.
-
MvNormal supports Cholesky Decomposition now for increased speed and numerical stability.
-
Many optimizations and speed-ups.
-
NUTS implementation now matches current Stan implementation.
-
Add higher-order integrators for HMC.
-
ADVI stopping criterion implemented.
-
Improved support for theano's floatX setting to enable GPU computations (work in progress).
-
MvNormal supports Cholesky Decomposition now for increased speed and numerical stability.
-
Added support for multidimensional minibatches
-
Added
Approximation
class and the ability to convert a sampled trace into an approximation via itsEmpirical
subclass. -
Model
can now be inherited from and act as a base class for user specified models (see pymc.models.linear). -
Add MvGaussianRandomWalk and MvStudentTRandomWalk distributions.
-
GLM models do not need a left-hand variable anymore.
-
Refactored HMC and NUTS for better readability.
-
Add support for Python 3.6.
-
Bound now works for discrete distributions as well.
-
Random sampling now returns the correct shape even for higher dimensional RVs.
-
Use theano Psi and GammaLn functions to enable GPU support for them.
We are proud and excited to release the first stable version of PyMC3, the product of more than 5 years of ongoing development and contributions from over 80 individuals. PyMC3 is a Python module for Bayesian modeling which focuses on modern Bayesian computational methods, primarily gradient-based (Hamiltonian) MCMC sampling and variational inference. Models are specified in Python, which allows for great flexibility. The main technological difference in PyMC3 relative to previous versions is the reliance on Theano for the computational backend, rather than on Fortran extensions.
Since the beta release last year, the following improvements have been implemented:
-
Added
variational
submodule, which features the automatic differentiation variational inference (ADVI) fitting method. Also supports mini-batch ADVI for large data sets. Much of this work was due to the efforts of Taku Yoshioka, and important guidance was provided by the Stan team (specifically Alp Kucukelbir and Daniel Lee). -
Added model checking utility functions, including leave-one-out (LOO) cross-validation, BPIC, WAIC, and DIC.
-
Implemented posterior predictive sampling (
sample_ppc
). -
Implemented auto-assignment of step methods by
sample
function. -
Enhanced IPython Notebook examples, featuring more complete narratives accompanying code.
-
Extensive debugging of NUTS sampler.
-
Updated documentation to reflect changes in code since beta.
-
Refactored test suite for better efficiency.
-
Added von Mises, zero-inflated negative binomial, and Lewandowski, Kurowicka and Joe (LKJ) distributions.
-
Adopted
joblib
for managing parallel computation of chains. -
Added contributor guidelines, contributor code of conduct and governance document.
- Argument order of tau and sd was switched for distributions of the normal family:
Normal()
LogNormal()
HalfNormal()
Old: Normal(name, mu, tau)
New: Normal(name, mu, sd)
(supplying keyword arguments is unaffected).
MvNormal
calling signature changed: Old:MvNormal(name, mu, tau)
New:MvNormal(name, mu, cov)
(supplying keyword arguments is unaffected).
We on the PyMC3 core team would like to thank everyone for contributing and now feel that this is ready for the big time. We look forward to hearing about all the cool stuff you use PyMC3 for, and look forward to continued development on the package.
The following authors contributed to this release:
Chris Fonnesbeck [email protected] John Salvatier [email protected] Thomas Wiecki [email protected] Colin Carroll [email protected] Maxim Kochurov [email protected] Taku Yoshioka [email protected] Peadar Coyle (springcoil) [email protected] Austin Rochford [email protected] Osvaldo Martin [email protected] Shashank Shekhar [email protected]
In addition, the following community members contributed to this release:
A Kuz [email protected] A. Flaxman [email protected] Abraham Flaxman [email protected] Alexey Goldin [email protected] Anand Patil [email protected] Andrea Zonca [email protected] Andreas Klostermann [email protected] Andres Asensio Ramos Andrew Clegg [email protected] Anjum48 Benjamin Edwards [email protected] Boris Avdeev [email protected] Brian Naughton [email protected] Byron Smith Chad Heyne [email protected] Corey Farwell [email protected] David Huard [email protected] David Stück [email protected] DeliciousHair [email protected] Dustin Tran Eigenblutwurst [email protected] Gideon Wulfsohn [email protected] Gil Raphaelli [email protected] Gogs [email protected] Ilan Man Imri Sofer [email protected] Jake Biesinger [email protected] James Webber [email protected] John McDonnell [email protected] Jon Sedar [email protected] Jordi Diaz Jordi Warmenhoven [email protected] Karlson Pfannschmidt [email protected] Kyle Bishop [email protected] Kyle Meyer [email protected] Lin Xiao Mack Sweeney [email protected] Matthew Emmett [email protected] Michael Gallaspy [email protected] Nick [email protected] Osvaldo Martin [email protected] Patricio Benavente [email protected] Raymond Roberts Rodrigo Benenson [email protected] Sergei Lebedev [email protected] Skipper Seabold [email protected] Thomas Kluyver [email protected] Tobias Knuth [email protected] Volodymyr Kazantsev Wes McKinney [email protected] Zach Ploskey [email protected] akuz [email protected] brandon willard [email protected] dstuck [email protected] ingmarschuster [email protected] jan-matthis [email protected] jason JasonTam22@gmailcom kiudee [email protected] maahnman [email protected] macgyver [email protected] mwibrow [email protected] olafSmits [email protected] paul sorenson [email protected] redst4r [email protected] santon [email protected] sgenoud [email protected] stonebig Tal Yarkoni [email protected] x2apps [email protected] zenourn [email protected]
Probabilistic programming allows for flexible specification of Bayesian statistical models in code. PyMC3 is a new, open-source probabilistic programmer framework with an intuitive, readable and concise, yet powerful, syntax that is close to the natural notation statisticians use to describe models. It features next-generation fitting techniques, such as the No U-Turn Sampler, that allow fitting complex models with thousands of parameters without specialized knowledge of fitting algorithms.
PyMC3 has recently seen rapid development. With the addition of two new major features: automatic transforms and missing value imputation, PyMC3 has become ready for wider use. PyMC3 is now refined enough that adding features is easy, so we don't expect adding features in the future will require drastic changes. It has also become user friendly enough for a broader audience. Automatic transformations mean NUTS and find_MAP work with less effort, and friendly error messages mean its easy to diagnose problems with your model.
Thus, Thomas, Chris and I are pleased to announce that PyMC3 is now in Beta.
- Transforms now automatically applied to constrained distributions
- Transforms now specified with a
transform=
argument on Distributions.model.TransformedVar
is gone. - Transparent missing value imputation support added with MaskedArrays or pandas.DataFrame NaNs.
- Bad default values now ignored
- Profile theano functions using
model.profile(model.logpt)
- A. Flaxman [email protected]
- Andrea Zonca [email protected]
- Andreas Klostermann [email protected]
- Andrew Clegg [email protected]
- AustinRochford [email protected]
- Benjamin Edwards [email protected]
- Brian Naughton [email protected]
- Chad Heyne [email protected]
- Chris Fonnesbeck [email protected]
- Corey Farwell [email protected]
- John Salvatier [email protected]
- Karlson Pfannschmidt [email protected]
- Kyle Bishop [email protected]
- Kyle Meyer [email protected]
- Mack Sweeney [email protected]
- Osvaldo Martin [email protected]
- Raymond Roberts [email protected]
- Rodrigo Benenson [email protected]
- Thomas Wiecki [email protected]
- Zach Ploskey [email protected]
- maahnman [email protected]
- paul sorenson [email protected]
- zenourn [email protected]