Re: 'Floating point exception'

On Wed, Mar 18, 2009 at 12:34 AM, Paolo <[email protected]> wrote:
>>  This is CRM114, version 20081111-BlameBarack, rev 4797 (TRE 0.7.5
>
> like previous one - indeed, I missed the 'rev 4797': means you already had
> GerH build installed.

And you just discovered a design bug in my revision process, thank you. Grrrrr.
Anyway, I'm a bit pressed for time these days (big SQL job going on
and some very special SSL alongside, so little time for crm114 right
now).

I've had fp errors happen in the past - I reviewed the code last year
and there were a few spots where things go ape when you would refute
as many documents as you've trained (repeated refuting of one document
is counted as rounds: after all, the stat counters in the CSS dbs
count rounds, not docs). I have tested that scenario following that
review for a few classifiers and fixed the code - or so I believe -
but this aspect hasn't been tested very thoroughly yet.

Then there's also some changes to cut down on the epsilon noise in the
code, which was really bad for my test suites, but then that might
have introduced a new hotspot for floating point value underrun. Sh**.

I hope you can find me that floating point error line (and a wee bit
of context: "It's in this multiMB logfile here" is good enough for
me).


The alternative (don't know how 'happy' this is going to make you) is
to ditch the CSS files, make sure you don't do <refute> training and
then pump the test set through from the start. (If you've prepped your
test corpus a little by now ;-)) then you should have metadata to go
with each article which would permit a full-auto classify-and-train
process for the whole shebang.  But that's me being nasty - I know how
long it took me to set up test suites like that for myself. Stumble
and fall was (and is) the order of the day.)


Back to SQL server hacking after one more reply...



>
> So I wonder if 'trace' mode was enough to make the exception disappear, or
> if it wasn't written to the logfile.

:-S   I've seen weirder stuff happen...

Ehm, extra info to be had:

1)
full output of
crm -v

(GerH spits out a serious number of lines there; espeically the latest
as it 'lists' the grammar rudimentarily)


2)
just as a reminder for when I need to review my code: which classifier
are you using exactly? (plus attributes, such as <unigram>, etc.,
thanks!)
choices are (but not limited to): OSB, OSBF, Hyperspace, markovian






-- 
Met vriendelijke groeten / Best regards,

Ger Hobbelt

--------------------------------------------------
web:    http://www.hobbelt.com/
        http://www.hebbut.net/
mail:   [email protected]
mobile: +31-6-11 120 978
--------------------------------------------------

------------------------------------------------------------------------------
Apps built with the Adobe(R) Flex(R) framework and Flex Builder(TM) are
powering Web 2.0 with engaging, cross-platform capabilities. Quickly and
easily build your RIAs with Flex Builder, the Eclipse(TM)based development
software that enables intelligent coding and step-through debugging.
Download the free 60 day trial. http://p.sf.net/sfu/www-adobe-com