Stochastic ICA-AROMA bug

[effigies]

• https://circleci.com/gh/effigies/fmriprep/589#tests/containers/1
• https://circleci.com/gh/poldracklab/fmriprep/929#tests/containers/1

I suspect this is related to either the low resolution or the small number of time points. It's the ICA-AROMA interface failing to run, so the solution might need to be in nipype.

[jdkent]

The problem file is here in feature_scores.txt

nan 6.564704524321367529e-01 2.187500000000000000e-01 3.239780910722724755e-02
nan 5.158288285626534408e-01 8.854166666666665741e-02 2.391670704351533949e-02
nan 4.715828885449906416e-01 8.854166666666665741e-02 4.360936751167206887e-02
nan 6.033063945444272491e-01 2.187500000000000000e-01 1.898491122542641033e-02
nan 5.930657172866073967e-01 8.697916666666666297e-01 3.077832062697738458e-02

The first column corresponds to maxRPcorr. I'll keep looking to see what going wrong.

[effigies]

I wonder if this line should be using np.nanmean instead.

[jdkent]

I made a python notebook for the function in question and used the melodic_mix and motion parameters from the failing subject. The number of NaNs is small so I believe your intuition is correct, we can just ignore the iterations that didn't work and average over the correct ones. I'll submit a pull request to ICA-AROMA.

Good Catch!