Commits on Feb 27, 2018

1. Remove 'thresh' option in fitwarp2d. …

   Singular-value decomposition of the basis vectors can return
   singular values that differ by 8--10 orders of magnitude. The
   default value of thresh (1e-5) would discard several of these smaller
   values.  That's a problem because the _inverses_ of the singular
   values are actually used to produce the solution in _svd. Discarding
   the largest singular values didn't seem to help either. Since it
   wasn't entirely clear why some of the singular values were being
   discarded, I commented out that line of code, and the associated
   documentation, so now we keep all of the singular values.
   
   I also added a detailed explanation of what fitwarp2d, _svd, and
   _mkbasis actually do.  And of course, some tests.
   
   If somebody with a lot of linear algebra experience is convinced
   of a a good reason to discard some of these singular values, we
   can revisit this (hence why I haven't removed the code, just
   commented it out for now).

   

   d-lamb committed Feb 27, 2018
   Copy the full SHA

   4ff7c4f View commit details

   Browse the repository at this point in the history

2. Try out a fix for memory freeing issue in warp2d. …

   There seems to be a problem with threads and Windows and freeing
   memory.  Previously the warp2d test would crash on Strawberry Perl
   with a "free to wrong pool" error.
   
   I followed the advice from BrowserUK in this PerlMonks thread:
   http://www.perlmonks.org/?node_id=742205 to create a pass-through
   routine to free the interolation kernel's memory.  Thanks for asking
   the similar question 9 years ago, Rob!
   
   These changes don't cause any tests to fail on my Mac, we shall see
   on the AppVeyor.

   

   d-lamb committed Feb 27, 2018
   Copy the full SHA

   8fe8cb5 View commit details

   Browse the repository at this point in the history