Development Changelog

This changelog only lists rather macroscopic changes to PyMVPA. The full VCS changelog for 2.x series of PyMVPA is available here:

Note

You could find relevant information on 0.4.x series at http://v04.pymvpa.org .

In addition there is also a somewhat unconventional visual changelog.

‘Closes’ and ‘Fixes’ statement IDs refer to the Debian and Github bug tracking systems accordingly and can be queried by visiting the URLs:

http://bugs.debian.org/<closed bug id>

https://github.com/PyMVPA/PyMVPA/issues/<fixed bug id>

Releases

  • 2.6.1 (Fri, Apr 14 2017)

    Primarily a bugfix release

    • Fixes

      • Deprecation: fixed typo in manhattan_distance
      • Documentation: various outdated links and project names fixed
      • Removed spurious quote within command line invocation of prep-afni-surf
      • Fixed computation of the edge lengths on the surface, which could have resulted in slightly incorrect surface searchlight sizes

    • New functionality

      • CDist for cross-validated dissimilarity measures computation

    • Enhancements

      • Hyperalignment - output_dim option could be used to perform dimensionality reduction

        via SVD while composing the common space

        • nproc and joblib_backend parameters could be used to parallelize the last step of the alignment

  • 2.6.0 (Sat, 26 Aug 2016)

    Primarily a bugfix release with some added new functionality. People contributed code to the release: Matteo Visconti dOC, Yaroslav Halchenko, Nikolaas N. Oosterhof, Feilong Ma, Christopher J. Markiewicz, Swaroop Guntupalli.

    • IMPORTANT possibly backward-incompatible fixes
      • Dataset’s vstack() and hstack() now by default would drop those feature or sample (fa, sa) attributes which do not have the same value across all datasets (“drop_nonunique”). Previous behaviour was to update aggregated collections, so the attribute value of the last dataset would have been stored in the stacked dataset. Such behaviour could be brought back by fa="update" for vstack or similar value for sa="update" for hstack calls. If you find that some sa/fa you have relied on using in your code is no longer available after stacking, verify that you did intend to maintain the “last known” value, and adjust argument in stacking function to “update”.
    • Fixes
      • Fixed minor bug in computing ico linear divisions while working with surfaces
      • Handling of ref_ds in SearchlightHyperalignment
      • Compatibility fixes for scipy 0.18.0 and nibabel 2.1.0.
    • New functionality
      • pymvpa2 scatter command line and scatter module to scatter plot datasets and nifti volumes, with coloring based on spatial location (see e.g. OHBM12 poster for an example, proper demo is coming)
    • Enhancements
      • Allow for “4D” mri mask volumes with degenerate time dimension (e.g. coming from AFNI)
      • pymvpa2 ttest could operate now on h5save’d datasets
      • It is possible now to h5save trained Hyperalignment instances
      • Balancer and AttributePermutator now got rng argument to seed RNG. Please use an int as a seed specification if you want random selections/permutations be consistent across searchlights

  • 2.5.0 (Thu, 12 May 2016)

    • Fixes
      • Various python3 related small fixes
      • Minor fix allowing adhoc searchlights (e.g. gnbsearchlight) to work with CustomPartitioner
      • Fixed SmartVersion to not infinitely loop upon receiving an awkward version string
    • New functionality
    • Enhancements
      • The ofmotionqc command line command has been renamed to plotmotionqc. It no longer requires a dataset formatted in OpenFMRI-style, but works with any organization of input data
      • SplitRFE docstring example reordered suggested mappers (l2 -> abs -> mean)
      • Show only summary of representation of internal _src2aux of VolumeMaskDictionary object, which otherwise could be too big to print

  • 2.4.3 (Tue, 15 Mar 2016)

    • Fixes
      • Fixed memory leaking in libsvm bindings introduced in 2.4.2 (gh-443)
      • Avoid overriding parameters defined in subclasses with the ones from super class
      • Address freshish deprecations (warnings) in sklearn, scipy, etc
      • Workaround in the test for numpy’s corrcoef issue resulting in corrcoef a bit outside of [-1, 1] range

  • 2.4.2 (Tue, 8 Mar 2016)

    • Fixes
      • Important: Reverse mapping of some chained Flatten/StaticSelection mappers did not work correctly e.g. if you selected some features from already masked fmri_dataset. This could have resulted in incorrect cluster counts by GroupClusterThreshold. Please recreate your datasets and re-estimate GroupClusterThreshold if that was the case for you
      • ad-hoc searchlights (gnb, m1nn) can now operate with partitioners which leave some samples out of training and testing sets. Also splitter argument was added to them for greater flexibility
      • Due to the bug in OpenfMRI datasets’ TR within NIfTIs being hardcoded to wrong 1.0, so scan_key.txt will now be consulted if TR is 1 in the .nii*
      • Compatibility with numpy 1.10 fixes
      • CachedQueryEngine acquired .ids making it compatible with some ad-hoc searchlights
      • FeatureSelection acquired __iadd__ fixing the incorrect behavior upon reverse after a sequence of feature selections
    • Deprecations/removal
      • Hamster is gone.
    • Enhancements
      • Bundled version of libsvm updated to 3.12. Now includes maxiter setting which prevents infinite looping which can happen in some rare cases
      • A swarm of stylistic improvements (“is not”, PEP8, etc) which should not affect functionality but could result in more robust operation
      • CrossValidation can now operate with a None generator (i.e. partitioner) using solely Splitter to generate a single split on original dataset. Provides easier means for “cross-classification”
      • M1NNSearchlight can now do classification based on correlation distance (just provide dfx=one_minus_correlation to kNN)
      • libsvm bindings for SVM were refactored to interface via svmc not _svmc interface, which made them also compatible with swig 3.x
      • to_npz() and from_npz() to interface Datasets through NumPy’s npz files
      • Variety of PEP8 and other tune ups for more readable code
      • SplitRFE can now work with static measures (e.g. OneWayAnova) and BinaryFxFeaturewiseMeasure. So do feature selection with nested cross-validation without double-dipping!
    • New functionality
      • FactorialPartitioner for factorial designs to cross-validate across sub-ordinate category samples (more efficient/avoids previously recommend ChainMapper of NFoldPartitioner and Sifter)

  • 2.4.1 (Wed, 18 Nov 2015)

    • New functionality
      • gifti can write GIFTI files that contain both dataset samples and surface anatomy (vertices and faces). Such GIFTI files can be read by FreeSurfer.
    • Deprecations/removal
      • tools/niils – tool removed, since the functionality was moved into nibabel under the name nib-ls
      • Drop support for nibabel < 2.0.0
    • Enhancements
      • “Native” use of duecredit to provide citations for PyMVPA itself and functionality/methods it implements.
      • Unified use of os.path.join as pathjoin.
      • ProcrusteanMapper computes reconstruction now more efficiently (just a transpose with proper scaling) in case of non-oblique transformations.
    • Refactorings/misc changes
      • ProcrusteanMapper now just returns transpose in reverse if transformation is non-oblique (instead of an explicit inverse).
    • Fixes
      • 2.4.0 was released with incorrect __version__ (as 2.3.1)
      • Fixes to ofmotionqc command implementation
      • Variety of fixes for compatibility with recent matplotlib, python3
      • Fixes to SVDMapper in reverse when projection is not a matrix

  • 2.4.0 (Mon, 11 May 2015)

    • New functionality
      • Support for CoSMoMVPA (http://cosmomvpa.org) in cosmo providing dataset input/output (cosmo_dataset() and map2cosmo()) and neighborhood input (CosmoQueryEngine). This allows for for running searchlights (mvpa2.datasets.cosmo.CosmoSearchlight) on data from CoSMoMVPA (fMRI and MEEG).
      • remove_nonfinite_features() removes features with non-finite values, i.e. NaNs or Infs, for any sample.
      • binomial_proportion_ci() for computing confidence intervals on proportions of Bernoulli trial outcomes.
      • New mapper for removing sample means from features: MeanRemoval
      • New algorithm for statistical evaluation of clusters in accuracy maps of group-based searchlight classification analyses. This is essentially an improved implementation of Stelzer et al., NeuroImage, 2013: GroupClusterThreshold
      • New identity mapper. Does nothing, but goes were only mappers can go.
      • Simplified selection of samples/feature in a dataset. One can now specify sets of attribute values to define sample/feature subsets.
      • IO adaptor for OpenFMRI-formated datasets: OpenFMRIDataset. Load arbitrary bits from such a dataset, or automatically build event-related dataset (optionally with NiPy-based HRF-modeling). tutorial_data_25mm was converted to OpenFMRI layout and extended also with 1slice flavor.
      • New command line command to generate a motion plot for an OpenFMRI-formated dataset.
      • New convenience functions for boxplots and outlier detection.
      • Reincarnated (similar functionality was removed for 2.0 release) convenience methods ( match() and select()) to ease selecting parts of a dataset
      • New command line component ttest to compute a simple T-test across dataset samples.
    • Enhancements
      • ProductFlattenMapper accepts explicit names of factors in the constructor.
      • HollowSphere() can now, optionally, include the center feature.
      • fmri_dataset() no longer stores original copy of the NIfTI file header – it converts it to dict representation to remain portable. Use strip_nibabel() to convert old datasets to new format if/when necessary.
      • Datasets created with fmri_dataset() do no longer contain instances of NiBabel classes. This improves robustness of re-loading serialized datasets in environments with changing NiBabel version. All previous functionality remains available. Dataset now contain an additional imgaffine attribute with the affine transformation of the source image.
    • Fixes
      • Hyperalignment with regularization (alpha != 1.0) was producing incorrect transformations because they were driven by offsets of the last subject. Fixed by not “auto_train”ing regularization projection.
      • plot_lightbox() should take a slice index from the last dimension, not the leading one if no slices argument was provided.
      • Improved Python3k compatibility in state, tests, and stats modules, and in libsvmlrc msvc building.
      • Partial fix for compatibility with ancient scipy on SPARC using cosmo.
      • Various fixes for example scripts and documentation snippets.
      • SKLTransformer can now deal with transform() implementations that do not support a y argument.

  • 2.3.1 (Tue, 20 May 2014)

    Primarily a bugfix release pushed out to avoid mvpa2.suite meltdown if new scipy 1.4.0 is used.

    • API changes
      • Deprecation: Parameter now uses constraints argument of type Constraint instead of string allowedtype. allowedtype argument will be removed completely in the future 2.4 release.
    • New changes
      • dummies now provides utterly useful RandomClassifier and others for code testing which could also be used to verify absent double-dipping etc.
    • Enhancements
      • FxMapper now will provide consistent order of groups of items. It also got a new argument order with available value of ‘occurrence’ to that groups would get ordered by their occurance in the original dataset.
    • Fixes
      • CorrStability should be able to deal with other sample attributes (not only ‘targets’) and should divide by variance correctly to provide correlation coefficient as output.
      • robustify check scipy’s rdist which should avoid crash upon import of mvpa2.suite because of stripped down scipy 1.4.0 API.
      • various typos in docstrings (we do welcome contributions ;) ).

  • 2.3.0 (Thu, 5 March 2014)

    • Warning: Due to a significant number of new features and some internal changes, loading of HDF5 files saved with previous versions might be impaired.
    • New functionality (>73 commits)
    • Enhancements (>130 commits)
      • fmri_dataset() load support for AFNI NIFTI volumes with time in the fifth dimension (data is automagically squeezed to 4D).
      • map2nifti() sets cal_ fields to correspond to the range of the data.
      • Many parts of the tutorial were reworked, set of examples expanded, and also converted to IPython notebooks.
      • eventrelated() extended to output GLM regression fits as features from HRF models (relying on NiPy for GLM modeling).
      • Class parameters now support programatic validation of values, auto-generated documentation and improved error messages.
      • More informative progress bar for long running processes, such as searchlights.
      • Replace all (broken) implementations for similarity structure analyses with new measures for computing pairwise pattern distances, and their consistency, or similarity to a target structure (rsa).
      • New examples to show integration with scikit-learn implementations of classification, regression, and data transformation algorithm.
    • Fixes (>88 commits)
      • Makefile fetch-data retrieves the data correctly.
      • HDF5 backend fixes to deal with nested/recursive structures and higher tolerance in loading HDF5 from older version.

  • 2.2.0 (Sun, Sep 16 2012)

    • New functionality (14 commits)
    • New HDF5-based storage backend for Searchlight the can significantly speed up serialization of large result dataset in parallelized computations.
    • New fast searchlight M1NNSearchlight (and helper sphere_m1nnsearchlight()) to run mean-1-nearest-neighbor searchlights.
    • New mappers for adding an axis to a dataset (AddAxisMapper), and for transposing a dataset (TransposeMapper).
    • Improved implementation of SciPy’s ttest_1samp() with support for masked arrays and alternative hypotheses.
    • Individual tutorial chapters are now available for download as IPython notebooks. A rst2ipynb converter is available in tools/.
    • New pymvpa2-tutorial command line utility to start a PyMVPA tutorial session, either in a console IPython session, or using the IPython notebook server.
    • New wrapper functions for data generators/loaders in sklearn.datasets, available in mvpa2.datasets.sources.sklearn_data.
    • Enhancements (89 commits)
      • Initial Python 3 compatibility (spear-headed by Tiziano Zito).
      • Bayesian hypothesis testing with BayesConfusionHypothesis now supports literal hypotheses specification, custom hypotheses subsets, and computing of posterior probabilities.
      • Allow for accessing fitted distributions in MCNullDist.
      • Extensions and improvements to the tutorial chapter on statistical evaluation.
      • Expose distance function as a property dfx of kNN.
      • Extended Sifter with ability to discard unbalanced partitions.
      • h5save() now creates missing directories automatically by default.
      • Dedicated training for Hyperalignment, and new auto-train capability.
      • BayesConfusionHypothesis now computes optional posterior probabilities, and supports hypothesis definitions using literal labels.
    • API changes
      • All command line tools have been renamed to have a consistent ‘pymvpa2-‘ prefix.
    • Fixes (77 commits)
      • HDF5 now properly stores object-type ndarray, where it the array shape was unintentionally modified on-load before (Fixes #84).
      • HDF5 can now reconstruct ‘builtin’ objects (Fixes #86).
      • Check value data type and convert to float when collecting performance statistics to avoid numerical problems.
      • Do not fail in BayesConfusionHypothesis when a dataset does not provide class labels.

  • 2.1.0 (Fri, June 29 2012)

    • Fixes
      • mask2slice() failed to convert an array of False values into slice(None, 0, None) (Fixes #56).
      • A number of fixes to the HDF5 IO code that ignored parts of an object’s state when custom __reduce__() implementations were used (Fixes #42), and had problems storing metaclass types (Fixes #78).
      • Proper single quotes in documentation code snippets within PDFs.
      • Memory leak (model pointer) in LIBSVM bindings.
    • Enhancements
      • All searchlight implementations can now optionally store the IDs of all features for each generated ROI (conditional attr. roi_feature_ids)
      • Add scatter_neighborhoods() to aid sparse sampling of spaces.
      • Add ConfusionMatrixError to compute confusion matrices with an error function interface (e.g. for CrossValidation(errorfx=...)). This class existed for a long time, but was hidden in the unit tests.
      • Add Confusion to compute confusion matrices with a Node interface (e.g. for CrossValidation(postproc=...)). This is useful if confusion matrices are necessary as an intermediate result and further processing with other nodes is desired.
    • New functionality
      • Add BayesConfusionHypothesis to perform Bayesian hypothesis testing of multi-class confusion statistics. This is useful to assess the likelihood of a particular (or all possible) grouping of classes being distinguishable.
      • Add FxyMapper to perform arbitrary computations involving two datasets.
      • Add CombinedMapper to run a dataset through a set of mappers and combine their outputs.
      • Add UnivariateStatsModels a wrapper for using models from the statsmodels package as a FeaturewiseMeasure.
      • Add dCOV and dcorcoef() to quantify independence of (multivariate) signals.
    • API changes
      • Deprecating GLM that is now implemented with UnivariateStatsModels. This deprecated GLM class no longer supports the zstat calculation, and none of its previous conditional attributes are available anymore.

  • 2.0.1 (Tue, Mar 27 2012)

    Primarily a bugfix release

    • Fixes (21 BF commits)
      • HDF5 storage – handle loading of objects with bound builtin_function_or_method.
      • Use system-wide autosummary/generate.py for sphinx >= 1.1.2 (Closes: #658593).
      • ConditionalAttribute should not loose value in deepcopy when default off (Fixes #63).
      • Correct handling of scalar mean/std values for ZScoreMapper.
      • MRI data import via _img2data now works with unicode filenames (Fixes #60).
      • Should work with IPython >= 0.11 now (Fixes #59).
      • Various small fixes to improve tests and functionality.
      • Fix SMLR segfaults on Windows (Thanks cgohlke for the patch).
    • Enhancements (29 ENH, OPT, and NF commits)
      • FxMapper calls functions natively (instead of a slow row/column at a time) if they carry axis as the 2nd argument. Provides tremendous speed up for mean_sample() etc.
      • xrandom_unique_combinations() generator for random unique combinations.
      • dual_gaussian() made more robust by not handling negative values for the standard deviations.
      • dual_positive_gaussian().
      • Expose “sensitivities” for PLR.
      • NFoldPartitioner float option for cvtype and intelligent behavior on ‘random’ limited by ‘count’ given a large number of folds.
      • Add few additional learners from sklearn to the warehouses: ExtraTrees, RandomForest, LassoLarsIC.
      • __repr__ for :class:`Partitioner`s.
      • Add new performance metric – F1 score – in the confusion matrix summary stats.
      • CachedQueryEngine does not rely now on id but on the exact value of the query parameters (converted to hashable types). Before it could behave incorrectly on rare occasions.
    • API changes

  • 2.0.0 (Mon, Dec 19 2011)

    This release aggregates all the changes occurred between official releases in 0.4 series and various snapshot releases (in 0.5 and 0.6 series). To get better overview of high level changes see release notes for 0.5 and 0.6 as well as summaries of release candidates below

    • Fixes (23 BF commits)
      • Significance level in the right tail was fixed to include the value tested – otherwise resulted in optimistic bias (or absurdly high significance in improbable case if all estimates having the same value).
      • Compatible with the upcoming IPython 0.12 and renamed sklearn (Fixes #57).
      • Do not double-train slave classifiers while assessing sensitivities (Fixes #53).
    • Enhancements (30 ENH + 3 NF commits)
      • Resolving voting ties in kNN based on mean distance, and randomly in SMLR.
      • kNN‘s ca.estimates now contains dictionaries with votes for each class.
      • Consistent zscoring in Hyperalignment.

  • 2.0.0~rc5 (Wed, Oct 19 2011)

    • Major: to allow easy co-existence of stable PyMVPA 0.4.x, 0.6 development mvpa module was renamed into mod:mvpa2.
    • Fixes
      • Compatible with the new Shogun 1.x series.
      • Compatible with the new h5py 2.x series.
      • mvpa-prep-fmri – various compatibility fixes and smoke testing.
      • Deepcopying SummaryStatistics during __add__.
    • Enhancements
      • Tutorial uses mvpa2.tutorial_suite now.
      • Better suppression of R warnings when needed.
      • Internal attributes of many classes were exposed as properties.
      • More unification of __repr__ for many classes.

  • 0.6.0~rc4 (Wed, Jun 14 2011)

    • Fixes
      • Finished transition to nibabel conventions in plot_lightbox.
      • Addressed matplotlib.hist API change.
      • Various adjustments in the tests batteries (nibabel 1.1.0 compatibility, etc)
    • New functionality
      • Explicit new argument flatten to from_wizard – default behavior changed if mapper was provided as well
    • Enhancements
      • Elaborated __str__ and __repr__ for some Classifiers and Measures

  • 0.6.0~rc3 (Thu, Apr 12 2011)

    • Fixes
      • Bugfixes regarding the interaction of FlattenMapper and BoxcarMapper that affected event-related analyses.
      • Splitter now handles attribute value None for splitting properly.
      • GNBSearchlight handling of roi_ids.
      • More robust detection of mod:scikits.learn and nipy externals.
    • New functionality
      • Added a Repeater node to yield a dataset multiple times and Sifter node to exclude some datasets. Consequently, the “nosplitting” mode of Splitter got removed at the same time.
      • tools/niils – little tool to list details (dimensionality, scaling, etc) of the files in nibabel-supported formats.
    • Enhancements
      • Numerous documentation fixes.
      • Various improvements and increased flexibility of null distribution estimation of Measures.
      • All attribute are now reported in sorted order when printing a dataset.
      • fmri_dataset now also stores the input image type.
      • Crossvalidation can now take a custom Splitter instance. Moreover, the default splitter of CrossValidation is more robust in terms of number and type of created splits for common usage patterns (i.e. together with partitioners).
      • CrossValidation takes any custom Node as errorfx argument.
      • ConfusionMatrix can now be used as an errorfx in Crossvalidation.
      • LOE(ACC): Linear Order Effect in ACC was added to ConfusionMatrix to detect trends in performances across splits.
      • A Nodes postproc is now accessible as a property.
      • RepeatedMeasure has a new ‘concat_as’ argument that allows results to be concatenated along the feature axis. The default behavior, stacking as multiple samples, is unchanged.
      • Searchlight now has the ability to mark the center/seed of an ROI in with a feature attribute in the generated datasets.
      • debug takes args parameter for delayed string comprehensions. It should reduce run-time impact of debug() calls in regular, non -O mode of Python operation.
      • String summaries and representations (provided by __str__ and __repr__) were made more exhaustive and more coherent. Additional properties to access initial constructor arguments were added to variety of classes.
    • Internal changes
      • New debug target STDOUT to allow attaching metrics (e.g. traceback, timestamps) to regular output printed to stdout
      • New set of decorators to help with unittests
        • @nodebug to disable specific debug targets for the duration of the test.
        • @reseed_rng to guarantee consistent random data given initial seeding.
        • @with_tempfile to provide a tempfile name which would get removed upon completion (test success or failure)
      • Dropping daily testing of maint/0.5 branch – RIP.
      • Collections were provided with adequate (deep|)copy. And Dataset was refactored to use Collections copy method.
      • update-* Makefile rules automatically should fast-forward corresponding website-updates branch
      • MVPA_TESTS_VERBOSITY controls also numpy warnings now.
      • Dataset.__array__ provides original array instead of copy (unless dtype is provided)

    Also adapts changes from 0.4.6 and 0.4.7 (see corresponding changelogs).

  • 0.6.0~rc2 (Thu, Mar 3 2011)

    • Various fixes in the mvpa.atlas module.

  • 0.6.0~rc1 (Thu, Feb 24 2011)

  • 0.5.0 (sometime in March 2010)

    This is a special release, because it has never seen the general public. A summary of fundamental changes introduced in this development version can be seen in the release notes.

    Most notably, this version was to first to come with a comprehensive two-day workshop/tutorial.

  • 0.4.7 (Tue, Mar 07 2011) (Total: 12 commits)

    A bugfix release

    • Fixed
      • Addressed the issue with input NIfTI files having scl_ fields set: it could result in incorrect analyses and map2nifti-produced NIfTI files. Now input files account for scaling/offset if scl_ fields direct to do so. Moreover upon map2nifti, those fields get reset.
      • doc/examples/searchlight_minimal.py - best error is the minimal one
    • Enhancements
      • GNB can now tolerate training datasets with a single label
      • TreeClassifier can have trailing nodes with no classifier assigned

  • 0.4.6 (Tue, Feb 01 2011) (Total: 20 commits)

    A bugfix release

    • Fixed (few BF commits):
      • Compatibility with numpy 1.5.1 (histogram) and scipy 0.8.0 (workaround for a regression in legendre)
      • Compatibility with libsvm 3.0
      • PLR robustification
    • Enhancements
      • Enforce suppression of numpy warnings while running unittests. Also setting verbosity >= 3 enables all warnings (Python, NumPy, and PyMVPA)
      • doc/examples/nested_cv.py example (adopted from 0.5)
      • Introduced base class LearnerError for classifiers’ exceptions (adopted from 0.5)
      • Adjusted example data to live upto nibabel’s warranty of NIfTI standard-compliance
      • More robust operation of MC iterations – skip iterations where classifier experienced difficulties and raise an exception (e.g. due to degenerate data)

  • 0.4.5 (Fri, Oct 01 2010) (Total: 27 commits)

    A bugfix release

    • Fixed (13 BF commits):
      • Compatible with LIBSVM >= 2.91 (Closes: #583018)
      • No string exceptions raised (Python 2.6 compatibility)
      • Setting of shrinking parameter in sg interface
      • Deducing number of SVs for SVR (LIBSVM)
      • Correction of significance in the tails of non-parametric tests
    • Miscellaneous:

  • 0.4.4 (Mon, Feb 2 2010) (Total: 144 commits)

    Primarily a bugfix release, probably the last in 0.4 series since development for 0.5 release is leaping forward.

    • New functionality (19 NF commits):
      • GNB implements Gaussian Naïve Bayes Classifier.
      • read_fsl_design() to read FSL FEAT design.fsf files (Contributed by Russell A. Poldrack).
      • SequenceStats to provide basic statistics on labels sequence (counter-balancing, autocorrelation).
      • New exceptions DegenerateInputError and FailedToTrainError to be thrown by classifiers primarily during training/testing.
      • Debug target STATMC to report on progress of Monte-Carlo sampling (during permutation testing).
    • Refactored (15 RF commits):
      • To get users prepared to 0.5 release, internally and in some examples/documentation, access to states and parameters is done via corresponding collections, not from the top level object (e.g. clf.states.predictions instead of soon-to-be-deprecated clf.predictions). That should lead also to improved performance.
      • Adopted copy.py from python2.6 (support Ellipsis as well).
    • Fixed (38 BF commits):
      • GLM output does not depend on the enabled states any more.
      • Variety of docstrings fixed and/or improved.
      • Do not derive NaN scaling for SVM’s C whenever data is degenerate (lead to never finishing SVM training).
      • sg :
        • KRR is optional now – avoids crashing if KRR is not available.
        • tolerance to absent set_precompute_matrix in svmlight in recent shogun versions.
        • support for recent (present in 0.9.1) API change in exposing debug levels.
      • Python 2.4 compatibility issues: kNN and IFS

  • 0.4.3 (Sat, 5 Sep 2009) (Total: 165 commits)

    • Online documentation editor is no longer available due to low demand – please submit changes via email.
    • Performance (Contributed by Valentin Haenel) (3 OPT commits):
      • Further optimized LIBSVM bindings.
      • Copy-if-sorted in selectFeatures.
    • New functionality (25 NF commits):
      • ProcrusteanMapper with orthogonal and oblique transformations.
      • Ability to generate simple reports using reportlab. See/run examples/match_distribution.py for example.
      • TreeClassifier – construct simple hierarchies of classifiers.
      • wtf() to report information about the system/PyMVPA to be included in the bug reports.
      • Parameter ‘reverse’ to swap training/testing splits in Splitter .
      • Example code for the analysis of event-related dataset using ERNiftiDataset.
      • toEvents() to create lists of Event.
      • mvpa-prep-fmri was extended with plotting of motion correction parameters.
      • ColumnData can be explicitly told either file contains a header.
      • In XMLBasedAtlas (e.g. fsl atlases) it is now possible to provide custom ‘image_file’ to get maps or indexes for the areas given an atlas’s volume registered into subject space.
      • Updated included LIBSVM version to 2.89 and provided support for its “silencing”.
    • Refactored (27 RF commits):
      • Dataset’s copy() with deep=False allows for shallow copying the dataset.
      • FeatureSelectionClassifier s in warehouse not to reuse the same classifiers, but to use clones.
    • Fixed (70 BF commits):
      • OneWayAnova: previously degrees of freedom were not considered while computing F-scores.
      • Majority voting strategy in kNN: it was not working.
      • Various fixes to ensure cross-platform building (numpy header locations, etc).
      • Stability fixes in ConfusionMatrix.
      • idsonboundaries(): samples at the end of the sequence were not handled properly.
      • Proper “untraining” of FeatureSelectionClassifier s classifiers which use sensitivities: it could lead to various unpleasant side-effects if the same slave classifier was used simultaneously by multiple MetaClassifiers (like TreeClassifier).
    • Documentation (25 DOC commits): citations, spelling corrections, etc.

  • 0.4.2 (Mon, 25 May 2009)

    • New correlation stability measure (CorrStability).
    • New elastic net classifier (ENET).
    • New GLM-Net regression/classifier (GLMNET).
    • New measure CompoundOneWayAnova.
    • New measure DSMDatasetMeasure.
    • New meta-measure TScoredFeaturewiseMeasure.
    • New basic GLM implementation.
    • New examples for Gaussian process regression.
    • New example showing a searchlight analysis employing a dissimilarity matrix based measure.
    • New ZScoreMapper.
    • New import helper for FSL design matrices (FslGLMDesign).
    • New implementation of a mapper using a self-organizing map (SimpleSOMMapper) and a corresponding example.
    • Matplotlib backend is now configurable via MVPA_MATPLOTLIB_BACKEND.
    • PyMVPA version is now avialable from mvpa.__version__.
    • Renamed mvpa.misc.plot.errLinePLot to plotErrLine() for consistency.
    • Fixed NFoldSplitter to support N-3 and larger splits.
    • Improved speed of LIBSVM backend. Thanks to Valentin Haenel and Tiziano Zito.
    • Updated included LIBSVM version to 2.89.
    • Adjust LIBSVM Python interface for recent NumPy API and latest LIBSVM release 2.89.
    • Refactored examples parser into a standalone tool to turn PyMVPA examples into restructured text sources.

  • 0.4.1 (Sat, 24 Jan 2009)

    • Unit tests and example data are now also installed. In conjunction with mvpa.test(), this allow to easily run unittests from within Python.
    • NiftiDataset capable to handle files with less than 4 dimensions, which can, optionally, be provided as a list of filenames or NiftiImage objects. That makes it easy to load data from a sequence of files.
    • Changes (code refactorings) which might impact any user who imports from suite:
      • Pre-populated warehouses of classifiers and regressions are renamed from clfs and regrs into clfswh and regrswh respectively.
      • Hamster is not derived from dict any longer – just from a basic object class. API includes methods ‘dump’, ‘asdict’ and a property ‘registered’.
    • Changes (code refactorings) which should not impact any user who imports from suite:
      • Meta classifiers definitions moved from base into meta.
      • Splitters definitions moved from splitter into splitters

  • 0.4.0 (Sat, 15 Nov 2008)

    • Add Hamster, as a simple facility to easily store any serializable objects in a compressed file and later on resurrect all of them with a single line of code.
    • SVM backend is now configurable via MVPA_SVM_BACKEND (libsvm or shogun).
    • Non-deterministic tests in the unittest battery are now configurable via MVPA_TESTS_LABILE.
    • New helper to determine and plot the best matching distribution(s) for the data (matchDistribution, plotDistributionMatches). It is WiP thus API can change in the upcoming release.
    • Simplifies API of mappers.
    • Splitters can now limit the number of splits automatically.
    • New CombinedMapper to map between multiple, independent dataspace and a common feature space.
    • New ChainMapper to create chains of mappers of abitrary lenght (e.g. to build preprocessing pipelines).
    • New EventDataset to rapidly extract boxcar-shaped samples from data array using a simple list of Event definitions.
    • Removed obsolete MetricMapper class. Mapper itself provides the facilities for dealing with metrics.
    • BoxcarMapper can now handle data with more than four dimensions/axis and also performs reverse mapping of single boxcar samples.
    • FslEV3 can now convert EV3 files into a list of Event instances.
    • Results of tests for external dependencies are now stored in PyMVPA’s config manager (mvpa.cfg) and can be stored to a file (not done automatically at the moment). This will significantly decrease the time needed to import the mvpa module, as it prevents the repeated and lengthy tests for working externals.
    • Initial support for ROC computing and AUC as an accuracy measure.
    • Weights of LARS are now available via LARSWeights.
    • Added an initial list of MVPA-related references to the manual, tagged with keywords and comments as well is DOI or similar URL reference to the original document.
    • Added initial glossary to the manual.
    • New ‘Module reference’, as a middle-ground between manual and API reference.
    • New manual section about meta-classifiers (contributed by James M. Hughes).
    • New minimal example for a ‘getting started’ section in the manual.
    • Former MVPA_QUICKTEST was renamed to MVPA_TESTS_QUICK.
    • Update installation instructions for RPM-based distributions to make use of the OpenSUSE Build Service.
    • Updated install instructions for several RPM-based GNU/Linux distributions.
    • Switch from distutils to numpy.distutils (no change in dependencies).
    • Depend on PyNIfTI >= 0.20081017.1 and gain a smaller memory footprint when accessing NIfTI files via all datasets with NIfTI support.
    • Added workaround to make PyMVPA work with older Shogun releases and those from 0.6.4 on, which introduced backward-incompatible API changes.

  • 0.3.1 (Sun, 14 Sep 2008)

    • New manual section about feature selection with a focus on RFE. Contributed by James M. Hughes.
    • New dataset type ChannelDataset for data structured in channels. Might be useful for data modalities like EEG and MEG. This dataset includes support for common preprocessing steps like resampling and baseline signal substraction.
    • Plotting of topographies on heads. Thanks to Ingo Fründ for contributing this code. Additionally, a new example shows how to do such plots.
    • New general purpose function for generating barplots and candlestick plots with error bars (plotBars()).
    • Dataset supports mapping of string labels onto numerical labels, removing the need to perform this mapping manually in user code. ‘clfs_examples.py’ is adjusted accordingly to demonstrate the new feature.
    • New mvpa.clfs.base.Classifier.summary() method to dump classifier settings.
    • Improved and more flexible plotERPs().
    • New IterativeRelief sensitivity analyzer.
    • Added visualization of confusion matrices via mvpa.clfs.transerror.ConfusionMatrix.plot() inspired by Ingo Fründ.
    • The PyMVPA version is now globally available in mvpa.pymvpa_version.
    • BugFix: TuebingenMEG reader failed in some cases.
    • Several improvements (docs and implementation) for building PyMVPA on MacOS X.
    • New convenience accessor methods (select(), where() and __getitem__()) for :class`~mvpa.datasets.base.Dataset`.
    • New mvpa.seed() function to configure the random number generators from user code.
    • Added reader for a MEG sensor locations format (TuebingenMEGSensorLocations).
    • Initial model selection support for GRP (using openopt).
    • And tons of minor bugfixes, additional tests and improved documentation.

  • 0.3.0 (Mon, 18 Aug 2008)

    • Import of binary EEP files (used by EEProbe) and EEPDataset class.
    • Initial version of a meta dataset class (MetaDataset). This is a container for multiple datasets, which behaves like a dataset itself.
    • Regression performance is summarized now within RegressionStatistics.
    • Error functions: CorrErrorPFx, RelativeRMSErrorFx.
    • Measures: CorrCoef.
    • Data generators: chirp, wr1996
    • Few more examples: curvefitting, kerneldemo, smellit, projections
    • Updated kNN classifier. kNN is now able to use custom distance function to determine that nearest neighbors. It also (re)gained the ability to do simple majority or weighted voting.
    • Some initial convenience functions for plotting typical results and data exploration.
    • Unified configuration handling with support for user-specific and analysis-specific config files, as well as the ability to override all config settings via environment variables. The configuration handling is used for PyMVPA internal settings, but can also be easily used for custom (user-)settings.
    • Improved modularity, e.g. SciPy is not required anymore, but still very useful.
    • Initial implementations of ICA and PCA mapper using functionality provided by MDP. These mappers are more or less untested and should be used with great care.
    • Further improved docstrings of some classes, but still a long way to go.
    • New ‘boxcar’ mapper, which is the similar to the already present transformWithBoxCar() function, but implemented as a mapper.
    • New SampleGroupMapper that can be used for e.g. block averaging of samples. See new FAQ item.
    • Stripped redundant suffixes from module names, e.g. mvpa.datasets.niftidataset -> mvpa.datasets.nifti
    • mvpa.misc.cmdline variables opt* and opts* were groupped within opt and optss class instances. Also names of the options were changed to match ‘dest’ of the options. Use tools/refactor.py to quickly fix your custom code.
    • Change all references to PyMVPA website to www.pymvpa.org.
    • Make website stylesheet compatible with sphinx 0.4.
    • Several minor improvements of the compatibility with MacOS.
    • Extended FAQ section of the manual.
    • Bugfix: double_gamma_hrf() ignoring K2 argument.

  • 0.2.2 (Tue, 17 Jun 2008)

    • Extended build instructions: Added section on OpenSUSE.
    • Replaced ugly PYMVPA_LIBSVM environment variable to trigger compiling the LIBSVM wrapper with a proper ‘–with-libsvm’ switch in setup.py. Additionally, setup.py now detects if included LIBSVM has been built and enables LIBSVM wrapper automatically in this case.
    • Added proper Makefiles for LIBSVM copy, with configurable compiler flags.
    • Added ‘setup.cfg’ to remove the need to manually specify swig-opts (Windows specific configuration is in ‘setup.cfg.win’).

  • 0.2.1 (Sun, 15 Jun 2008)

    • Several improvements to make building PyMVPA on Windows systems easy (e.g. added dedicated Makefile.win to build a binary installer).
    • Improved and extended documentation for building and installing PyMVPA.
    • Include a minimal copy of the required (patched) LIBSVM library (currently version 2.85.0) for convenience. This copy is automatically compiled and used for the LIBSVM wrapper when PyMVPA built using the Make approach.

  • 0.2.0 (Wed, 29 May 2008)

    • New Splitter class (HalfSplitter) to split into first and second half.
    • New Splitter class (CustomSplitter) to allow for splits with an arbitrary number of datasets per split and the ability to specify the association of samples with any of those datasets (not just the validation set).
    • New sparse multinomial logistic regression (SMLR) classifier and associated sensitivity analyzer.
    • New least angle regression classifier (LARS).
    • New Gaussian process regression classifier (GPR).
    • Initial documentation on extending PyMVPA.
    • Switch to Sphinx for documentation handling.
    • New example comparing the performance of all classifiers on some artificial datasets.
    • New data mapper performing singular value decomposition (SVDMapper) and an example showing its usage.
    • More sophisticated data preprocessing: removal of non-linear trends and other arbitrary confounding regressors.
    • New Harvester class to feed data from arbitrary generators into multiple objects and store results of returned values and arbitrary properties.
    • Added documentation about how to build patched libsvm version with sane debug output.
    • libsvm bindings are not build by default anymore. Instructions on how to reenable them are available in the manual.
    • New wrapper from SVM implementation of the Shogun toolbox.
    • Important bugfix in RFE, which reported incorrect feature ids in some cases.
    • Added ability to compute stats/probabilities for all measures and transfer errors.

  • 0.1.0 (Wed, 20 Feb 2008)

    • First public release.