PyMC：专注于先进的马尔可夫链蒙特卡罗算法
PyMC：专注于先进的马尔可夫链蒙特卡罗算法
v3.7
New features
- 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.
Maintenance
- All occurances 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
.
Deprecations
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 pymc3-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.
Contributors sorted by number of commits
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
Assets
2
v3.6
Assets
2
This is a major new release from 3.5 with many new features and important bugfixes. The highlight is certainly our completely revamped website: https://docs.pymc.io/
Note also, that this release will be the last to be compatible with Python 2. Thanks to all contributors!
New features
- Track the model log-likelihood as a sampler stat for NUTS and HMC samplers
(accessible astrace.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).
Maintenance
- 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).
Deprecations
- Renamed
sample_ppc()
andsample_ppc_w()
tosample_posterior_predictive()
andsample_posterior_predictive_w()
, respectively.
v3.6.rc1
Assets
2
This will be the last Python 2 compatible release.
New features
- Track the model log-likelihood as a sampler stat for NUTS and HMC samplers
(accessible astrace.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).
Maintenance
- 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).
Deprecations
- Renamed
sample_ppc()
andsample_ppc_w()
tosample_posterior_predictive()
andsample_posterior_predictive_w()
, respectively.
v3.5
v3.5 Final
fonnesbeck released this
New features
- 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
Fixes
- 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
v3.4.1
v3.4.1 Final
fonnesbeck released this
There was no 3.4 release due to a naming issue on PyPI.
New features
- 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).
Fixes
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.
Deprecations
- 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.
v3.3
v3.3 Final
fonnesbeck released this
New features
- 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/pymc3/)
- 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)
Fixes
- 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)
Deprecations
- Old (
minibatch-
)advi
is removed (#2781)
热门度与活跃度
10.0
10.0
Watchers：249 |
Star：4684 |
Fork：1133 |
创建时间： 2009-05-05 17:43:50 |
最后Commits： 昨天 |
许可协议：View license |
ca35f4f
ColCarroll released this
Nov 30, 2019
· 2 commits to master since this release
New features
DifferentialEquation
. See #3590 and #3634.Data
andDeterministic
variables when graphing models with graphviz. PR #3491.Matern12
covariance function for Gaussian processes. This is the Matern kernel with nu=1/2.@
operator now works with random variables and deterministics #3619.Maintenance
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 #3508DEMetropolis
) is now set via thecores
argument. (#3559)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 #3535OrderedLogistic.__init__
that unnecessarily increased the dimensionality of the underlyingp
. Related to issue issue #3535 but was not the true cause of it.pm.sample
now it should be called usingpm.sample_smc
3579scaling
andtune_scaling
arguments as is a better idea to always allow SMC to automatically compute the scaling factor 3625multiprocessong
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.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 #3535OrderedLogistic.__init__
that unnecessarily increased the dimensionality of the underlyingp
. Related to issue issue #3535 but was not the true cause of it.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 3553theano.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)