mvpa2.datasets.cosmo

Dataset from CoSMoMVPA

This module provides basic I/O support for datasets in CoSMoMVPA. The current implementation provides (1) loading and saving a CoSMoMVPA dataset struct, which is converted to a PyMVPA Dataset object; and (2) loading a CoSMoMVPA neighborhood struct, which is converted to a CoSMoQueryEngine object that inherits from QueryEngineInterface.

A use case is running searchlights on MEEG data, e.g.:

1. FieldTrip is used to preprocess MEEG data
2. CoSMoMVPA is used to convert the preprocessed MEEG data to a CoSMoMVPA dataset struct and generate neighborhood information for the searchlight
3. this module (mvpa2.datasets.cosmo) is used to import the preprocessed MEEG data and the neighborhood information into PyMVPA-objects
4. PyMVPA is used to run a searchlight with a Measure of interest
5. this module (mvpa2.datasets.cosmo) is used to export the searchlight output to a CoSMoMVPA dataset struct
6. CoSMoMVPA is used to convert the CoSMoMVPA dataset struct with the searchlight output to a FieldTrip struct
7. FieldTrip is used to visualize the results

Example

Suppose that in Matlab using CoSMoMVPA, two structs were created:

ds (with fields .samples, .sa, .fa and .a)

containing a dataset (e.g. an fMRI, MEEG, or surface-based dataset). Such a struct is typically defined in CoSMoMVPA using cosmo_{fmri,meeg,surface}_dataset

nbrhood (with fields .neighbors, .fa and .a)

containing neighborhood information for each feature in ds. Such a struct is typically defined in CoSMoMVPA using cosmo_neighborhood.

Alternatively they can be defined in Matlab directly without use of CoSMoMVPA functionality. For a toy example, consider the following Matlab code:

>> ds=struct();
>> ds.samples=[1 2 3; 4 5 6];
>> ds.a.name='input';
>> ds.fa.i=[1 2 3];
>> ds.fa.j=[1 2 2];
>> ds.sa.chunks=[2 2]';
>> ds.sa.targets=[1 2]';
>> ds.sa.labels={'a','b','c','d';'e','f','g','h'};
>> save('ds_tiny.mat','-struct','ds');

>> nbrhood=struct();
>> nbrhood.neighbors={1, [1 3], [1 2 3], [2 2]};
>> nbrhood.fa.k=[4 3 2 1];
>> nbrhood.a.name='output';
>> save('nbrhood_tiny.mat','-struct','nbrhood');

These can be stored in Matlab by:

>> save('ds.mat','-struct','ds')
>> save('nbrhood.mat','-struct','nbrhood')

and loaded in Python using:

>>> import mvpa2
>>> import os
>>> from mvpa2.datasets.cosmo import from_any, CosmoSearchlight
>>> from mvpa2.mappers.fx import mean_feature
>>> data_path=os.path.join(mvpa2.pymvpa_dataroot,'cosmo')
>>> fn_mat_ds=os.path.join(data_path,'ds_tiny.mat')
>>> fn_mat_nbrhood=os.path.join(data_path,'nbrhood_tiny.mat')
>>> ds=from_any(fn_mat_ds)
>>> print ds

<Dataset: 2x3@float64, <sa: chunks,labels,targets>, <fa: i,j>, <a: name>> >>> qe=from_any(fn_mat_nbrhood) >>> print qe CosmoQueryEngine(4 center ids (0 .. 3), <fa: k>, <a: name>

where ds is a Dataset and qe a CosmoQueryEngine.

A Measure of choice can be used for a searchlight; here the measure simply takes the mean over features in each searchlight:: >>> measure=mean_feature()

A searchlight can be run the CosmoQueryEngine >>> sl=CosmoSearchlight(measure, qe) >>> ds_sl=sl(ds) >>> print ds_sl <Dataset: 2x4@float64, <sa: chunks,labels,targets>, <fa: k>, <a: name>>

Note that the output dataset has the feature and sample attributes taken from the queryengine, not the dataset.

Alternatively it is possible to run the searchlight directly using the filename of the neighborhood .mat file:

>>> sl=CosmoSearchlight(measure, fn_mat_nbrhood)
>>> ds_sl=sl(ds)
>>> print ds_sl

<Dataset: 2x4@float64, <sa: chunks,labels,targets>, <fa: k>, <a: name>>

which gives the same result as above.

Leaving the doctest format here, subsequently the result can be stored in Python using:

>> map2cosmo(ds_sl,'ds_sl.mat')

and loaded in Matlab using:

>> ds_sl=importdata('ds_sl.mat')

so that in Matlab ds_sl is a dataset struct with the output of applying measure to the neighborhoods defined in nbrhood.

Notes

• This function does not provide or deal with mappers associated with a dataset. For this reason map2nifti does not work on PyMVPA fmri datasets that were imported from CoSMoMVPA using this module. Instead, CoSMoMVPA’s map2fmri in Matlab can be used to map results to nifti and other formats
• The main difference between the searchlight approach in CoSMoMVPA versus PyMVPA is that CoSMoMVPA allows for setting feature (.fa) and dataset (.a) attributes for the output explicitly in its QueryEngineInterface. Use cases are (a) surface-based searchlight of fMRI data (with .fa set to the node indices of the output) and (b) MEEG searchlight of timelocked data that uses all (or a subset of the) sensors for the input and provides a time-course of MVP results for the output; where the input data has features of time by sensor, while the output data has only time.

Functions

cosmo_dataset(cosmo)	Construct Dataset from CoSMoMVPA format
from_any(x)	Load CoSMoMVPA dataset or neighborhood
map2cosmo(ds[, filename])	Convert PyMVPA Dataset to CoSMoMVPA struct saveable by scipy’s savemat

Classes

CosmoQueryEngine(mapping[, a, fa])	queryengine for neighborhoods defined in CoSMoMVPA.
CosmoSearchlight(datameasure, nbrhood[, ...])	Implement a standard Saerchlight measure, but with a separate postprocessing step that involves setting feature (.fa) and dataset (.a) attributes after the searchlight call has been made.