Re: GerH source distro at hebbut.net: HOW TO compile/compare/merge for Bill et al

On Wed, Apr 2, 2008 at 4:42 PM, Bill Y <[email protected]> wrote:
>    Neural Net.
>    See timings on full and reduced test set in megatest elsewhere. NN
>    takes it all. At least on 64-bit. ;-) Sorry.
>
>  Don't be sorry,  that's good to know.
>
>  Try running the "alternating_example_neural.crm" test separately (it's
>  standalone .crm code), with -t on.  Remember to delete i_test.css
>  and q_test.css first.
>
>  You should see a cascade of letters crawling up your screen
>  just like in The Matrix.  Darker letters (more ink) imply
>  sections of the neural net that got things worse; blanks
>  are sections that are within the learning tolerance of
>  being perfect.
>
>  In the first section, at around line 36, you should start to
>  see some streaks of white space; those should win out by line
>  63.  The respin (different random number initialization) starts
>  to streak at line 97 and then has a single long tail that finally
>  converges at iteration 200.
>
>  The whole thing takes only about 32 seconds to run on a 1.6 GHz
>  32-bit Linux laptop.
>
>  As a test- can you tell your compiler to generate 32-bit code
>  and then run the 32-bit code?
>
>  This is incompatibility on a _massive_ scale.

I have run that crm file before and yes, I saw The Matrix :-) Very nice view.

Thanks for this clarification; I'll have another look at it soon
enough, though right now my 1st priority regarding crm114 is doing
something illicit to mailreaver+mailtrainer.
Once that is working at a satisfactory level - I find I learn quite a
few new things about the crm language I didn't when I focussed on
megatest + my own stuff before :-) - I'll pay some very close
attention to NN.

BTW, did you check out the diff explanation I put on-line? some NN
stuff in there as well and I don't feel like writing that up again.
;-)


Regarding 64/32:

I have the Linux box for UNIX 64-bit development. The true 32-bit
development is done on the Windows machine. My (old?) copy of NN is
slow on both, but I can provide you with numbers and everything when
NN is the focus of my attention lateron. It'll _very_ probably take a
couple more days though :-(


BTW: expect compatibility issue at the same scale in CLUMP - yes I
know it's experimental, but even experimental code shouldn't barf so
gloriously on a box that's a little different. I'll get back to that
later too, because my thought is this:
Experimental or not, when you are able to produce stable results for a
single test (minor or no diff for megatest_knowngood over several wget
versions), that same test _must_ produce the same result on a
different box. If it doesn't, it's not algorithm screwup, it's
implementation detail bugs like out-of-bounds access and implicit
wordsize assumptions that do you in. And _that_ sort of thing should
be fixed from day 1, because if it isn't, it means implementation
thought processes are flawed at a deep detail level. Those kind of
thought bugs are more visible when it's all still a bit rough and
people haven't had time yet to cover their ass using 'hacks' to 'make
it work' on theirs, meaning faster diagnostic+fix. I still recall our
BitEntropy dance, I never took up the courage to 'assault' CLUMP in
the same way (SVM and SKS also have some issues, but less so), and now
that Neural Net is there, to my eye it's the third time in a row I get
to see the same type of symptoms. Now that scares _me_ right out of my
pants, because it makes me feel like I'm heading for a total mess.

If you are willing, we can set up another one of those 'dances' in a
while and see how reproducable we can get NN for the given [mega]test
set. Not because its production ready yet, but because (a) megatest
output should be as it is, anywhere, and (b) this is like training.
Training for something you don't want to know about, if you're still
sane. ;-))
(My goal: if I can get you to recognize some of the portability issues
in regular code detail, I get a better deep detail QC on the work of
the interns right at Command Central, and _that_ means I get to spend
less effort in long term. I'd like _that_. Lazy is my middle name.
Even while I'm a workaholic.)


Take care,

Ger





-- 
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
--------------------------------------------------

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace