Crossnobis

[bpinsard]

I developed 2 interfaces for Cross-validated mahalanobis distance analysis for RSA. (see: Reliability of dissimilarity measures for multi-voxel pattern analysis. Walther et al. 2015)
One is a Measure, and the other is an optimised Searchlight.
It is not finished yet, and there lack tests and proper doc, but I wanted to know if there is interest to include this in PyMVPA, and if so what are the best design choice.

For example in the searchlight, I allow to pass a residuals dataset to _sl_call, in order to compute covariance for each split/searchlight (i need to add this in the Measure too) to perform multivariate normalization in an computationally optimised way (computing sparse covariance of whole dataset between features in same searchlights only). But I don't know what would be the best way to pass it trough call of the parent Searchlight class.

Any comment welcome.
Thanks

[coveralls]

Coverage decreased (-16.1%) to 65.124% when pulling bf70cce on bpinsard:crossnobis into e834243 on PyMVPA:master.

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (master@11503a9). Click here to learn what that means.
The diff coverage is 80%.

@@            Coverage Diff            @@
##             master     #524   +/-   ##
=========================================
  Coverage          ?   61.42%           
=========================================
  Files             ?      298           
  Lines             ?    33847           
  Branches          ?     5146           
=========================================
  Hits              ?    20791           
  Misses            ?    11836           
  Partials          ?     1220

Impacted Files	Coverage Δ
mvpa2/support/utils.py	62.79% <ø> (ø)
mvpa2/mappers/__init__.py	100% <ø> (ø)
mvpa2/support/nibabel/afni_niml.py	77.43% <100%> (ø)
mvpa2/misc/io/base.py	78.99% <75%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 11503a9...3c775d7. Read the comment docs.

[coveralls]

Coverage decreased (-16.03%) to 65.213% when pulling 7ee40b8 on bpinsard:crossnobis into e834243 on PyMVPA:master.

[yarikoptic]

I thought I have left comments but apparently didn't ;)

1. code without tests is (most likely) broken code
2. changes to other classes seems broke existing code

I felt that something like that doesn't require an additional class(es) and could be accomplished within existing constructs... I could be wrong though -- want to see some tests/old code working before paying closer attention.

Cheers!

[coveralls]

Coverage decreased (-16.1%) to 65.13% when pulling 713cd8f on bpinsard:crossnobis into e834243 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.9%) to 65.142% when pulling 73d5f58 on bpinsard:crossnobis into c2e2628 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.9%) to 65.142% when pulling d29ae79 on bpinsard:crossnobis into c2e2628 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.9%) to 65.127% when pulling 887fc56 on bpinsard:crossnobis into c2e2628 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.8%) to 65.21% when pulling 74761ac on bpinsard:crossnobis into c2e2628 on PyMVPA:master.

[yarikoptic]

@bpinsard could you please add at least a single simple integration test so we could see intended usage and help bringing things into a state not making other functionality break etc

[coveralls]

Coverage decreased (-14.9%) to 65.127% when pulling 70e4c93 on bpinsard:crossnobis into 56d8861 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.9%) to 65.127% when pulling 70e4c93 on bpinsard:crossnobis into 56d8861 on PyMVPA:master.

[coveralls]

Coverage decreased (-15.0%) to 65.094% when pulling 11503a9 on bpinsard:crossnobis into 56d8861 on PyMVPA:master.

[coveralls]

Coverage decreased (-14.9%) to 65.176% when pulling 3c775d7 on bpinsard:crossnobis into 96b0b1a on PyMVPA:master.