Add Corrmap

[jona-sassenhagen]

Implementation of Corrmap; website, journal paper.

Corrmap is a procedure for semi-automatic selection of maps, especially for selecting common ICs for a number of subjects. The user inputs a list of fitted ICAs (such as for all of the subjects of a study) and denotes a template IC (such as a blink IC). The Corrmap algorithm then determines the best-fitting ICs for all ICAs and adds them to the ICAs in a specified field.

Here is a gist.

The routine should basically work and be somewhat robust. I'm open to any feedback.
This is not a translation of the original MATLAB program, but an attempt of recreating the behaviour as described in the journal paper.

Currently, the routine only takes ICs, but in the long term, it could be expanded to select any sets of matching topographies, such as for basic microstate analysis.

I'm not sure how a test could be built as the algorithm requires multiple IC decompositions.

[agramfort]

@jona-sassenhagen thanks heaps for sharing. Have you worked with this code on a few datasets? My logic is that if it's still early, it's better to iterate quickly on this code on your end and merge it after it proved useful on a few datasets.

my 2c

[jona-sassenhagen]

Here is another dataset, this time 4 participants with 64 channels.

Fastica
AMICA

Good results, at least in the second case. It does not perform well in the first because the decompositions are not very good.
So this is also an example of how AMICA, compared to Fastica, can give very good decompositions without any prior data cleaning or parameter tuning, and how it avoids splitting components. (It is possible Fastica would be much better with e.g. rejections enabled, I have not tried to fine-tune Fastica.)

I will run another test soon.

[agramfort]

thanks heaps for sharing.

[dengemann]

@jona-sassenhagen this looks pretty cool! Let me know when you need some reviewing.

[jona-sassenhagen]

Another demonstration: full study of 20 subjects. http://nbviewer.ipython.org/gist/jona-sassenhagen/ec915a7e9abda1e2828b
Completely missed the blink IC for 1 subject and chose the wrong IC for another. Not bad.
Also includes example of corrected vs. uncorrected data.

@dengemann , can you indeed take a quick look at the code itself? No details for now, just tell me if it generally works with the way MNE is laid out.

[dengemann]

@jona-sassenhagen no reason to use line breaks here
Also note that we use @verbose decorater together with a verbose keyword.

[jona-sassenhagen]

I'm deleting the verbose_return thing, it's dumb.

[dengemann]

@jona-sassenhagen I would put it in ICA, not in bads.

[dengemann]

@jona-sassenhagen apart from my comments looks like a good start to me!

[agramfort]

I would use the term labels as what you're doing is labeling components. So ica.labels

also make sure you check pep8 compliance of your file.

And it would be really great to have at least a smoke test.

[jona-sassenhagen]

Especially if you repeatedly do the same analysis, run ICA with a fixed random state, and store the correct templates.

[dengemann]

Ok. Just wanted to make sure all we have here is needed.

[jona-sassenhagen]

Travis:

Test file downloading ... ERROR
======================================================================
ERROR: Test file downloading

Is this my fault?

[larsoner]

No, sometimes we get those failures

[jona-sassenhagen]

Can I restart Travis somehow? It's all urlib errors.

[larsoner]

You have to have commit rights on the repo to restart, but I restarted it for you

[jona-sassenhagen]

Thanks @Eric89GXL

[jona-sassenhagen]

Again URLError: and TimeoutError: [Errno 110] Connection timed out ...

[larsoner]

Just ignore those for now -- if those are the only errors, is this otherwise ready for merge? We might end up having to merge some red PRs today because of connectivity issues, once we make sure that the only problems are connectivity errors.

[dengemann]

@jona-sassenhagen @Eric89GXL this PR induces 9% of coverage loss in ica.py ...

[jona-sassenhagen]

Can we check somewhere what lines are not being tested? I'm at pretty sure corrmap isn't even 10% of the ica code 

On Sat, Apr 18, 2015 at 11:29 AM -0700, "Denis A. Engemann" notifications@github.com wrote:

@jona-sassenhagen @Eric89GXL this PR induces 9% of coverage loss in ica.py ...

—
Reply to this email directly or view it on GitHub.

[larsoner]

You can't trust the coveralls report because coveralls is only called when a build succeeds

[larsoner]

...so you have to check the changes locally. Can you take a look and compare master to your PR @jona-sassenhagen ?

[dengemann]

I checked locally

Nostests --with-coverage

On 18 Apr 2015, at 21:08, Eric Larson notifications@github.com wrote:

...so you have to check the changes locally. Can you take a look and compare master to your PR @jona-sassenhagen ?

—
Reply to this email directly or view it on GitHub.

[coveralls]

Coverage decreased (-0.2%) to 90.017% when pulling 877a4db on jona-sassenhagen:patch-3 into 4ae85b5 on mne-tools:master.

[jona-sassenhagen]

For some reason we had plot=False for all tests. Coverage is much better now. Basically the lines it's not covering right now are some of the exceptions raised by the various failure options.

I'm considering adding the ability to deal with heterogenous IC objects, e.g. when ICA is called on bad channels.

[jona-sassenhagen]

@dengemann like this?

[jona-sassenhagen]

Would be very cool to have ICA interpolation ... I'm playing around with _interpolate_bads_eeg, but I'm too bad at lin alg to really make it work.

[dengemann]

@jona-sassenhagen closing in favor of #1985.