[Eeglablist] NaN in AMICA

[Tyler Grummett]

Dear EEGLABers,


More to the issues I've been having with AMICA, I was wondering if there was a way to continue where you left off. Im currently running multiple instances of AMICA in a super computer, which is linux based, and an error I most often receive is "error writing to output stream". While this is likely the super computer's fault, I was wondering if it was at all possible to redesign the AMICA code (for future application) so that it can refer to the output variables and go from where it crashed previously.


I can fully understand if this is not possible. Just a thought.


Tyler


*************************

Tyler Grummett ( BBSc, BSc(Hons I))
PhD Candidate
Brain Signals Laboratory
Multimodal Recording Facility
Flinders University Tonsley Building
Rm 4.17
Ext 19573
________________________________
From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> on behalf of Tyler Grummett <tyler.grummett at flinders.edu.au>
Sent: Tuesday, 8 May 2018 12:42:51 PM
To: EEGLABLIST
Subject: [Eeglablist] NaN in AMICA


Dear EEGLABers (particularly Jason Palmer)


Ive been using EEGLAB and AMICA for many years now and theres been an issue I've been meaning to ask about for a while now. Sometimes when I run an AMICA, the 'LL' and 'nd' in the final iteration are NaNs. For example (last couple lines of output):


...

 iter  3993 lrate =  0.0000004768 LL =   0.2151813642 nd =  0.0066193118, D =   0.11548E+01  0.11548E+01  ( 44.53 s, 198.0 h)
 iter  3994 lrate =  0.0000004768 LL =   0.2151813668 nd =  0.0066193124, D =   0.11548E+01  0.11548E+01  ( 44.42 s, 197.5 h)
 iter  3995 lrate =  0.0000004768 LL =   0.2151813695 nd =  0.0066193130, D =   0.11548E+01  0.11548E+01  ( 44.42 s, 197.5 h)
 iter  3996 lrate =  0.0000004768 LL =   0.2151813720 nd =  0.0066193133, D =   0.11548E+01  0.11548E+01  ( 44.56 s, 198.1 h)
 iter  3997 lrate =  0.0000004768 LL =   0.2151813746 nd =  0.0066193139, D =   0.11548E+01  0.11548E+01  ( 44.27 s, 196.8 h)
 iter  3998 lrate =  0.0000004768 LL =   0.2151813771 nd =  0.0066193143, D =   0.11548E+01  0.11548E+01  ( 44.55 s, 198.0 h)
 iter  3999 lrate =  0.0000004768 LL =   0.2151813797 nd =  0.0066193143, D =   0.11548E+01  0.11548E+01  ( 44.49 s, 197.7 h)
 iter  4000 lrate =  0.0000004768 LL =   0.2151813823 nd =  0.0066193152, D =   0.11548E+01  0.11548E+01  ( 44.05 s, 195.8 h)
 iter  4001 lrate =  0.0000004768 LL =   0.2151813824 nd =  0.0066193255, D =   0.11548E+01  0.11548E+01  ( 48.06 s, 213.6 h)
 iter  4002 lrate =  0.0000004768 LL =   0.2151813847 nd =  0.0066193309, D =   0.11548E+01  0.11548E+01  ( 44.39 s, 197.3 h)
 iter  4003 lrate =  0.0000004768 LL =   0.2151813871 nd =  0.0066193324, D =   0.11548E+01  0.11548E+01  ( 44.35 s, 197.1 h)
 iter  4004 lrate =  0.0000004768 LL =   0.2151813892 nd =  0.0066193335, D =   0.11548E+01  0.11548E+01  ( 44.49 s, 197.7 h)
 iter  4005 lrate =  0.0000004768 LL =   0.2151813914 nd =  0.0066193340, D =   0.11548E+01  0.11548E+01  ( 44.62 s, 198.2 h)
 iter  4006 lrate =  0.0000004768 LL =            NaN nd =           NaN, D =   0.11548E+01  0.11548E+01  ( 44.28 s, 196.7 h)
 Got NaN! Exiting ...
... done. Execution time: -10.23 h


Since first noticing this till now, I have code that checks the output for this issue and reruns the AMICA. A lot of the time it will eventually and properly end with the learning rate not decreasing by 1e-9 for five iterations (I think thats what it is). Ive checked the data, it seems fine.


I guess what I am asking is: "Is this an issue?" or am I incorrectly classifying it as an issue. If it is an issue, Im wondering why it doesnt reset itself.


Kind regards,

Tyler


*************************

Tyler Grummett ( BBSc, BSc(Hons I))
PhD Candidate
Brain Signals Laboratory
Multimodal Recording Facility
Flinders University Tonsley Building
Rm 4.17
Ext 19573
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20180605/b6d67a02/attachment.html>