Re: Function definition locations

I would love to start using the new sources, but I understand the
installer is still broken.  So I've been waiting to hear it's fixed.
Any news?  I've had enough headaches installing without having to
fight with removing, too.

On Apr 29, 4:02 pm, Kee Siong Ng <keesiong...@...> wrote:
> Buddha, you really need to be looking at the updated SVM code from
> the master branch of MADlib:
>
>    https://github.com/madlib/madlib/tree/master/methods/kernel-machines/
>
> First of all, the existing v0.1alpha versions are *wrong* in their
> margin update formulas; the errors are in the Kivinen et al papers,
> from where I took the algorithms.
>
> Secondly, the latest code are in C, which is *significantly* faster.
>
> Thirdly, the latest code actually allows the user to supply his/her
> own kernel function, instead of fixing that to be a vector inner
> product.
>
> Fourthly, the latest code conforms to MADlib's recommended API, and we
> would appreciate your input on how usable that API is.
>
> Thanks.
>
> On Apr 30, 3:42 am, buddha_314 <buddha_...@...> wrote:
>
>
>

Updating Versions

Yes, I know I should be using the latest and greatest.  My experience
with the installer isn't so great.  So what is the recommended update
path?

Re: Updating Versions

Sorry for the trouble with the installer. I'm not qualified to comment
on the installer so will leave that to the others to answer.

In the mean time, if you only need SVM for now, can you try changing
Config.yml to only install kernel-machines? (Just comment out all the
other methods.)

If that doesn't work, come back to me and I'll send you the installer-
processed source files that you can load directly using psql.

If you want to stay with v0.1alpha, at the very least, you need to
change the following lines:

In $GPHOME/ext/python/lib/python2.6/site-packages/madlib/kernel-
machines/src/pg_gp/online_sv.sql,

1) in the online_sv_cl_update() function, change
            svs.rho := svs.rho + eta * (1 - nu);     to
svs.rho := svs.rho - eta * (1 - nu);
    and   svs.rho := svs.rho - eta * nu;              to
svs.rho := svs.rho + eta * nu;

    Lines 440 and 442 at
https://github.com/madlib/madlib/blob/v0.1alpha/methods/kernel-machines/src/pg_gp/online_sv.sql_in

2) do the same for the online_sv_nd_update() function.

    Lines 485 and 487 at
https://github.com/madlib/madlib/blob/v0.1alpha/methods/kernel-machines/src/pg_gp/online_sv.sql_in

Re: Re: Function definition locations

Who sez the installer is broken?

Sent from a telephone.

On May 2, 2011, at 7:22 PM, buddha_314 <buddha_314@...> wrote:

> I would love to start using the new sources, but I understand the
> installer is still broken.  So I've been waiting to hear it's fixed.
> Any news?  I've had enough headaches installing without having to
> fight with removing, too.
> 
> On Apr 29, 4:02 pm, Kee Siong Ng <keesiong...@...> wrote:
>> Buddha, you really need to be looking at the updated SVM code from
>> the master branch of MADlib:
>> 
>>    https://github.com/madlib/madlib/tree/master/methods/kernel-machines/
>> 
>> First of all, the existing v0.1alpha versions are *wrong* in their
>> margin update formulas; the errors are in the Kivinen et al papers,
>> from where I took the algorithms.
>> 
>> Secondly, the latest code are in C, which is *significantly* faster.
>> 
>> Thirdly, the latest code actually allows the user to supply his/her
>> own kernel function, instead of fixing that to be a vector inner
>> product.
>> 
>> Fourthly, the latest code conforms to MADlib's recommended API, and we
>> would appreciate your input on how usable that API is.
>> 

Re: install confused by multiple Config.yml files

I should point out that the solution is provided as part of this
thread:

http://groups.google.com/group/madlib-user-forum/browse_thread/thread/3e3647c4f0c6d1a0

Basically, I had to remove a few copies of Config.yml and re-install.
Just hit me again this morning.

On Apr 24, 5:08 pm, buddha_314 <buddha...@...> wrote:
> I can't find the reason yet, but some script appears to copy the
> Config.yml file into build/lib/madpy/  and /usr/local/greenplum-
> db-4.0.1.0/ext/python/lib/python2.6/site-packages/madpy/config/scripts/
> Config.yml  I can't seem to convince the installer to use the edited
> version for another database/user.  None of the uninstall scripts will
> roll it back so far as I can tell.
>
> So once you start installing the Config.yml file becomes moot.  This
> is not expected to me, since I can' back it all up.
>
> Doing an uninstall doesn't remove this.
>
> On Apr 24, 2:44 pm, buddha_314 <buddha...@...> wrote:
>
>
>
>
>
>
>
> > I also noticed that changing the database name in madpy/Config.yml was

Cannot find svec_eq_non_zero

Doing a re-installing with the latest sources in github.  I believe
this means I'm using the SVM and SVEC versions Kee-Siong mentioned.
Running madpack install gives me the error below.

I have installed madlib on a few different databases, but I believe I
has over-written the sources.

Traceback (most recent call last):
  File "/usr/local/bin/madpack", line 248, in <module>
    main(sys.argv[1:])
  File "/usr/local/bin/madpack", line 237, in main
    install_version(args.scriptdir[0], instv, args.configdir[0])
  File "/usr/local/bin/madpack", line 106, in install_version
    __install(scriptdir,None,v,conf_dir)
  File "/usr/local/bin/madpack", line 92, in __install
    m.migrate(mignumber=num, migversion=version)
  File "/Library/Python/2.6/site-packages/madpy/madpack/
migrations.py", line 422, in migrate
    m.forwards()
  File "/Library/Python/2.6/site-packages/madpy/config/scripts/
000_Master.py", line 1935, in forwards
    --!""")
  File "build/bdist.macosx-10.6-universal/egg/pgdb.py", line 259, in
execute
  File "build/bdist.macosx-10.6-universal/egg/pgdb.py", line 289, in
executemany
pg.DatabaseError: error 'ERROR:  could not find function
"svec_eq_non_zero" in file "/sw/opt/postgresql-9.0/lib/madlib/
gp_svec.so"

Re: Cannot find svec_eq_non_zero

Re: Cannot find svec_eq_non_zero

Is this all compiled locally, or is there already and gp_svec.so in
the github repository I can use?

On May 7, 8:26 am, Aleks <Aleksander.Gora...@...> wrote:
> Sounds like you are probably still using older
> /sw/opt/postgresql-9.0/lib/madlib/gp_svec.so
>
> So, when running your cleanup you can try to rm this dir:
> /sw/opt/postgresql-9.0/lib/madlib
>
> BTW: next week we will have the new installer in the Master branch of
> github, so these problems should be soon over and hopefully new ones will be
> limited.

--

-- 
You received this message because you are subscribed to the Google Groups "MADlib User Forum" group.
To post to this group, send email to madlib-user-forum@...
To unsubscribe from this group, send email to madlib-user-forum+unsubscribe <at> googlegroups.com.
For more options, visit this group at http://groups.google.com/group/madlib-user-forum?hl=en.

Re: Cannot find svec_eq_non_zero

At this time, everything gets built locally (when you run "madpack 
install"). When we get to beta release later this month you will be getting 
compiled code in the package and only DB actions will be done locally.

multinomial logistic regression

Does madlib provides a way to multinomial logistic regression. I can
see the binomial version in the modules.