hama-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Yexi Jiang <yexiji...@gmail.com>
Subject Re: [DISCUSS] Roadmap for 0.7.0
Date Sat, 27 Jul 2013 02:52:39 GMT
Actually, the issue HAMA-770 (https://issues.apache.org/jira/browse/HAMA-770)
is related to this.

Regards,
Yexi


2013/7/26 Tommaso Teofili <tommaso.teofili@gmail.com>

> Ah yes, sure, I think that's a good idea.
>
> Thanks,
> Tommaso
>
>
> 2013/7/26 Yexi Jiang <yexijiang@gmail.com>
>
> > I mean we can use a generic model internally for MLP, linear regression,
> > logistic regression and so on. From the user's perspective, they use
> these
> > model separately (For example, three different classes). But in fact,
> these
> > classes are just the wrapper of the internal model.
> >
> >
> > 2013/7/26 Tommaso Teofili <tommaso.teofili@gmail.com>
> >
> > > I'm not sure I understand your question Yexi, would you like to re use
> > > existing models? For what? Feedforward prediction in a neural network?
> > >
> > > Regards,
> > > Tommaso
> > >
> > >
> > > 2013/7/25 Yexi Jiang <yexijiang@gmail.com>
> > >
> > > > > yes, they just use the plain gradient descent BSP implementation
> but
> > > one
> > > > > could do both with a neural network (MLP) as well, however I'd like
> > to
> > > > have
> > > > > separate implementations which can address different scenarios.
> > > >
> > > > Yes, from the user perspective, they should be different model. But
> is
> > it
> > > > OK to reuse the same model (the artificial neural network, a more
> > generic
> > > > MLP) internally?
> > > >
> > > >
> > > > 2013/7/25 Tommaso Teofili <tommaso.teofili@gmail.com>
> > > >
> > > > > 2013/7/25 Yexi Jiang <yexijiang@gmail.com>
> > > > >
> > > > > > I also think about to remove the regression.CostFunction and
> create
> > > one
> > > > > in
> > > > > > ml.math. But it seems that the argument list is quite different
> > from
> > > > the
> > > > > > DoubleDoubleFunction defined in ml.math.
> > > > >
> > > > >
> > > > > yes, that's because it came before the MLP Function API and was
> > > designed
> > > > > just for the case of regression.
> > > > > So I'd be +1 in standardizing our interfaces, making some of the
> > > changes
> > > > > mentioned before and more if needed.
> > > > >
> > > > >
> > > > > > I think we need to figure out how
> > > > > > to replace the regression.CostFunction with existing ones.
> > Moreover,
> > > I
> > > > > > found both the linear regression and logistic regression use
> > gradient
> > > > > > descent. So they can be treated as a two layer MLP in fact.
> > > > > >
> > > > >
> > > > > yes, they just use the plain gradient descent BSP implementation
> but
> > > one
> > > > > could do both with a neural network (MLP) as well, however I'd like
> > to
> > > > have
> > > > > separate implementations which can address different scenarios.
> > > > >
> > > > > Regards,
> > > > > Tommaso
> > > > >
> > > > >
> > > > > >
> > > > > > Regards,
> > > > > > Yexi
> > > > > >
> > > > > >
> > > > > > 2013/7/25 Tommaso Teofili <tommaso.teofili@gmail.com>
> > > > > >
> > > > > > > Hi Yexi,
> > > > > > >
> > > > > > > sorry for the late reply, I think we can improve a bit
our ML
> API
> > > by
> > > > > > adding
> > > > > > > (for example) generics to Functions, move
> regression.CostFunction
> > > > > > interface
> > > > > > > to math or remove it as some functions in math are themselves
> > cost
> > > > > > > functions, change MLP package name from perception to
> perceptron,
> > > > this
> > > > > > kind
> > > > > > > of small things that may result (IMHO) in a more readable
and
> > user
> > > > > > friendly
> > > > > > > API.
> > > > > > >
> > > > > > > My 2 cents,
> > > > > > > Tommaso
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > 2013/7/21 Yexi Jiang <yexijiang@gmail.com>
> > > > > > >
> > > > > > > > Hi, Tommaso,
> > > > > > > >
> > > > > > > > For the machine learning module, what kind of refactoring
do
> > you
> > > > > think
> > > > > > is
> > > > > > > > necessary?
> > > > > > > >
> > > > > > > > Regards,
> > > > > > > > Yexi
> > > > > > > >
> > > > > > > >
> > > > > > > > 2013/7/21 Edward J. Yoon <edwardyoon@apache.org>
> > > > > > > >
> > > > > > > > > Additionally, Queue is also one of big issues.
> > > > > > > > >
> > > > > > > > > On Sun, Jul 21, 2013 at 8:55 PM, Tommaso Teofili
> > > > > > > > > <tommaso.teofili@gmail.com> wrote:
> > > > > > > > > > Hi Edward,
> > > > > > > > > >
> > > > > > > > > > I'm still quite unsure about the status
of FT so it may
> be
> > > > worth
> > > > > > > doing
> > > > > > > > > some
> > > > > > > > > > work to make sure that is fully working
(but it may be
> just
> > > > me).
> > > > > > > > > > Also vertex storage in graph package should
be improved.
> > > > > > > > > > Then I'd say some refactoring of machine
learning module
> > APIs
> > > > > > > together
> > > > > > > > > with
> > > > > > > > > > addition of "Collaborative Filtering"  (and
eventually
> some
> > > > other
> > > > > > > > > > algorithms, but I'm still unsure there).
> > > > > > > > > >
> > > > > > > > > > My 2 cents,
> > > > > > > > > > Tommaso
> > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > 2013/7/19 Edward J. Yoon <edwardyoon@apache.org>
> > > > > > > > > >
> > > > > > > > > >> Hi all,
> > > > > > > > > >>
> > > > > > > > > >> Once HAMA-742 is done, users will be
able to install a
> > Hama
> > > > > > cluster
> > > > > > > on
> > > > > > > > > >> existing Hadoop 1.x and new Hadoop 2.x
without issues.
> > > > > > > > > >>
> > > > > > > > > >> I think urgent tasks are finished, now
it's time to
> > discuss
> > > > > about
> > > > > > > the
> > > > > > > > > >> future roadmap Hama 0.7 and begin enhancement
work.
> > > > > > > > > >>
> > > > > > > > > >> Please feel free to voice your opinions.
> > > > > > > > > >>
> > > > > > > > > >> Thanks.
> > > > > > > > > >>
> > > > > > > > > >> --
> > > > > > > > > >> Best Regards, Edward J. Yoon
> > > > > > > > > >> @eddieyoon
> > > > > > > > > >>
> > > > > > > > >
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > --
> > > > > > > > > Best Regards, Edward J. Yoon
> > > > > > > > > @eddieyoon
> > > > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > > > --
> > > > > > > > ------
> > > > > > > > Yexi Jiang,
> > > > > > > > ECS 251,  yjian004@cs.fiu.edu
> > > > > > > > School of Computer and Information Science,
> > > > > > > > Florida International University
> > > > > > > > Homepage: http://users.cis.fiu.edu/~yjian004/
> > > > > > > >
> > > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > --
> > > > > > ------
> > > > > > Yexi Jiang,
> > > > > > ECS 251,  yjian004@cs.fiu.edu
> > > > > > School of Computer and Information Science,
> > > > > > Florida International University
> > > > > > Homepage: http://users.cis.fiu.edu/~yjian004/
> > > > > >
> > > > >
> > > >
> > > >
> > > >
> > > > --
> > > > ------
> > > > Yexi Jiang,
> > > > ECS 251,  yjian004@cs.fiu.edu
> > > > School of Computer and Information Science,
> > > > Florida International University
> > > > Homepage: http://users.cis.fiu.edu/~yjian004/
> > > >
> > >
> >
> >
> >
> > --
> > ------
> > Yexi Jiang,
> > ECS 251,  yjian004@cs.fiu.edu
> > School of Computer and Information Science,
> > Florida International University
> > Homepage: http://users.cis.fiu.edu/~yjian004/
> >
>



-- 
------
Yexi Jiang,
ECS 251,  yjian004@cs.fiu.edu
School of Computer and Information Science,
Florida International University
Homepage: http://users.cis.fiu.edu/~yjian004/

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message