Release Notes – PyMVPA 0.5

For The Impatient

  • Datasets are no longer relatively static objects, but become flexible multi-purpose containers that can handle attributes for samples, feature, or whole datasets. There is some inital support for other datatypes than NumPy’s ndarrays, e.g. sparse matrices.

Critical API Changes

  • states -> ca (for conditional attributes). All attributes stored in collections (parameters for Classifiers in .params, states in .ca) should be accessed not at top level of the object but through a collection.
  • Dataset: behaves more like a NumPy array. No specialized Dataset classes, but constructors
    • MaskedDataset -> dataset_wizard
    • NiftiDataset -> fmri_dataset
    • ERNiftiDataset -> fmri_dataset + eventrelated_dataset (see event-related analysis example)
  • MRI volumes: 3,4D volumes (and coordinates) are exposed with following order of axes: t,x,y,z. Previously we followed a convention of t,z,y,x order of axis in volume data (to be consistent with PyNIfTI).
  • Masks (mask_mapper)
  • now [1,1,0] is not the same as [True, True, False]
  • We have weird (but consistent) conventions now - classes are CamelCased - factory functions (even for whatever might have been before a class) are in pythonic_style
  • detrend -> poly_detrend
  • perchunk=bool (in zscore/detrend) got refactored into chunks_attr=None or string to specify on which sample attribute to operate.
  • internally and as provided by mvpa2.suite, numpy is imported as np, and pylab is imported as pl

General Changes

Datasets

Sparse data support

Dataset in principal now support non-ndarray types for dataset samples. However, most parts of PyMVPA still assume an (at least) ndarray-like interface.

Splitters

  • permute -> permute_attr, so if you had permute=True, use attr='targets' if you like to permute targets