GitLab

Rather than replicating these in MATLAB, we now run the Python-based
postprocessing routines for PROFUMO. This makes things considerably
simpler, especially if we want to add features to postprocessing.

Now print the full error message before continuing.

The aim of this argument is to initialise with a better split between PFMs /
global noise. Interest is whether it is problematic when no global noise
exists.

But robust to new / missing methods.

Hopefully helpful for working out provenance of old results.

Just more consistent with the way things are normally presented.

New built-in `histogram` function is pretty decent.

Removes another hidden dependency.

Too many tests! This simplifies things, and means we don't need to worry about
changing to the external `dual_regression` interface.

As per the convention for external calls to PROFUMO/MELODIC too.

Keeps everything consistent, and relative performances easier to interpret.

Testing what effect this has on methods - performance on amplitudes is
currently very similar.

I.e. the correction proposed in the NAF2 paper.

This is not always max is best (e.g. the NAF2 plots), so we now explicitly plot
the optimal value too.

Now computes the proper `melodic` & `dual_regression` pipeline, which is a good
sanity check.

Make sure the NIfTI is at least a size that we can save.

In essence, this commit moves some of the hard coded paths out of the helper
functions and into `RunTests.m`, where things are a bit easier to track.

Add within subject (i.e. to complement across subject).

Now calculated relative to the ground truth interaction, just to make sure the
baseline is a sensible one.

I.e. the NAF2 stroy from Bijsterbosch et al., eLife, 2019.

The mixing can induce amplitude changes, so we renormalise here before
thresholding (and then add explicit variability later).

Hasn't been failing, but makes consistent with MELODIC behaviour.

Just adds annoying noise when editing with different editors, so fix all files
at once...

For some reason, not all versions of MATLAB propagate the zero maps through the
`pinv()`. Therefore have to re-null the timecourses etc.

 + Add variability in amplitudes to simulations.
 + Test for this as part of scoring procedures.
 + Load inferred PROFUMO amplitudes.

Just makes it a bit harder :-)