[petsc-dev] -snes_check_jacobian

[Jed Brown]

Options for drawing the result, using coloring instead of dense
differencing, and thresholding using rtol and atol to avoid false
positives. IIRC, the reason I added all of those variants was because I was
given some Jacobian code that was only incorrect when something occurred at
a boundary, and was not always obvious for small sizes. In any case,
drawing the difference is quite handy.


On Thu, Dec 13, 2012 at 3:29 PM, Peter Brune <prbrune at gmail.com> wrote:

> Reaction after looking at Barry's changeset -- yeah this is great.  Jed,
> does -compare_explicit_* do anything that's missing here?
>
> - Peter
>
>
>
> On Thu, Dec 13, 2012 at 5:22 PM, Peter Brune <prbrune at gmail.com> wrote:
>
>>
>>
>>
>> On Thu, Dec 13, 2012 at 5:10 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>
>>>
>>> On Dec 13, 2012, at 5:04 PM, Peter Brune <prbrune at gmail.com> wrote:
>>>
>>> > Dangit Barry!  I already started a patch moving all Jed's
>>> -snes_compare_* garbage in SNESComputeJacobian() over to a
>>> SNESJacobianMonitor registration type deal where it has functions you can
>>> register to run after computing the Jacobian.  This will allow you to do
>>> all sorts of nice comparisons.
>>>
>>>   Dang it we don't need yet another registration thingy, I used the
>>> SNESSetUpdate() hook.
>>>
>>>
>>   Actually I don't mind if you do something better and remove mine (it is
>>> the story of my life) but don't remove mine until yours produces as pretty
>>> an output and has good names for the options and yes
>>> snes_compare_explicit is a terrible name. At least put jacobian in the
>>> option so I have some clue that it has something to do with the jacobian!
>>>
>>>
>> I agree, and this seems like a fine place for it to happen.  I'll have to
>> look at exactly what you did, but you could save a Jacobian computation if
>> you put it after the user Jacobian has been computed, which is why I was
>> putting it as hooks to be run at the end of SNESComputeJacobian().
>>
>> I want similar facilities for the objectives as well.   There's already
>> an FD test objective in place (used by SNESTEST to verify if you have
>> written a proper objective -> function -> jacobian set).  If we had it so
>> you could register multiple tests (even in the update slot) we could
>> provide a whole suite of potential useful output without much work.
>>
>> - Peter
>>
>>   Barry
>>>
>>> >
>>> > Dangit Jed!  -snes_compare_* is a seeerious mess and shouldn't be in
>>> SNESComputeJacobian().  Barry's recent holy war against string comparisons
>>> will find this some day and there will be a lot of explaining to do.
>>> >
>>> > - Peter
>>> >
>>> >
>>> > On Thu, Dec 13, 2012 at 4:59 PM, Jed Brown <jedbrown at mcs.anl.gov>
>>> wrote:
>>> > Surely you were aware of the -snes_compare_* options that can be
>>> activated in SNESComputeJacobian().
>>> >
>>> > Can we settle on one common way to do this testing and perhaps remove
>>> -snes_type test.
>>> >
>>> >
>>> > On Thu, Dec 13, 2012 at 2:55 PM, Barry Smith <bsmith at mcs.anl.gov>
>>> wrote:
>>> >
>>> >    I got sick of -snes_type test only checking the Jacobian at the
>>> initial solution and not being able to test later Jacobians so I added
>>> -snes_check_jacobian that allows the regular running of SNESSolve() (also
>>> inside TS) and compares every Jacobian computed with a finite difference
>>> version.
>>> >
>>> >   For example, (already found one bug in one example :-(). This may be
>>> the best thing I've added to PETSc in a decade.
>>> >
>>> >    Barry
>>> >
>>> >
>>> > Barrys-MacBook-Pro:advection-diffusion-reaction barrysmith$ ./ex1
>>> -snes_check_jacobian -snes_monitor  -ts_type arkimex | more
>>> >     0 SNES Function norm 2.182384017537e+00
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.68322e-09 = ||J - Jfd||//J|| 6.69028e-06  = ||J - Jfd||
>>> >     1 SNES Function norm 4.695397453383e-07
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       2.45109e-09 = ||J - Jfd||//J|| 9.74235e-06  = ||J - Jfd||
>>> >     2 SNES Function norm 4.020379099215e-13
>>> >     0 SNES Function norm 6.799718390407e-01
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.65487e-09 = ||J - Jfd||//J|| 6.5776e-06  = ||J - Jfd||
>>> >     1 SNES Function norm 4.558192078678e-08
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       2.93714e-09 = ||J - Jfd||//J|| 1.16742e-05  = ||J - Jfd||
>>> >     2 SNES Function norm 9.907308296870e-14
>>> >     0 SNES Function norm 1.000842117652e+00
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.24476e-09 = ||J - Jfd||//J|| 4.94755e-06  = ||J - Jfd||
>>> >     1 SNES Function norm 9.875108391772e-08
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.10818e-10 = ||J - Jfd||//J|| 4.40468e-07  = ||J - Jfd||
>>> >     2 SNES Function norm 7.677760150017e-14
>>> >     0 SNES Function norm 2.179048757949e+00
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.32338e-09 = ||J - Jfd||//J|| 5.27058e-07  = ||J - Jfd||
>>> >     1 SNES Function norm 4.625448863612e-05
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >         Run with -snes_check_jacobian_display to show difference of
>>> hand-coded and finite difference Jacobian.
>>> >       1.57582e-09 = ||J - Jfd||//J|| 6.27588e-07  = ||J - Jfd||
>>> >     2 SNES Function norm 4.391784647745e-14
>>> >     0 SNES Function norm 6.761591873515e-01
>>> >         Testing hand-coded Jacobian, if the ratio is O(1.e-8), the
>>> hand-coded Jacobian is probably correct.
>>> >
>>> >
>>> >
>>> >
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20121213/1ca7718e/attachment.html>