One of the issues of GPU is data move. You can't do that without some
knowledge that would depend on data size, algorithm pipeline and other
stuff.
IMHO GPU will be usable when CPU and GPU memories will be integrated
without move cost. Before, GPU will be hype without mainstream usage.
Cheers
Matthieu
Le 7 mai 2014 08:07, "Ralf Gunter" <ralfgun...@gmail.com> a écrit :
> For what it's worth, I'm interested in getting bohrium[1] (sucessor of
> distnumpy with GPU & cluster support) and sklearn to play along nicely
> to speedup some of the regression algorithms (especially gaussian
> process) because we have a decently sized cluster and can't compromise
> on accuracy. The documentation is a bit... lacking, but it shouldn't
> be too hard to do this.
>
> [1] -- http://bohrium.bitbucket.org/
>
> 2014-05-07 1:49 GMT-05:00 Andy <t3k...@gmail.com>:
> > Hi Michal.
> > Please direct all such questions to the sklearn mailing list or
> > stackoverflow.
> > I doubt there will be any integration of GPU computation into numpy in
> > the near future (or probably ever).
> > There is also no plan to integrate GPU acceleration into scikit-learn,
> > mostly because it introduces a lot of dependencies.
> >
> > If your problem is kernel SVMs, GPUs don't really help much anyhow. If
> > your dataset is large, I would suggest using the kernel approximation
> > module.
> > This example illustrates the use and the gains in speed:
> >
> http://scikit-learn.org/dev/auto_examples/plot_kernel_approximation.html#example-plot-kernel-approximation-py
> >
> > Best,
> > Andy
> >
> >
> > On 05/05/2014 11:20 PM, Michal Sourek wrote:
> >> Hi Andreas,
> >> with regards to your long track of Scikit-learn development,
> >> let me raise one issue from code-execution point of view.
> >>
> >> Scikit has many marvelous tools.
> >>
> >> My primary concern is related to Supervised Learning of SVM-based
> >> Classifier.
> >>
> >> In the initial coarse optimisation scan over a non-convex problem,
> >> a GridSearchCV() use on the real-world dataset
> >> has rather long code-execution times,
> >> that even off-peak shifts scheduled onto available CPU/Cores
> >> resource-pool,
> >> available over weekend DataCenter shifts,
> >> cannot handle.
> >>
> >> Would indeed appreciate your comments, ideas or directions about the
> >> possible scikit-learn acceleration strategies available or
> >> possible-in-principle in this context.
> >>
> >> Are there, upon your knowledge, any R&D activities running in this
> >> direction?
> >>
> >> Have not found any promising candidates anywhere near the Scikit-learn
> >> framework. Also not aware about any industry-group spin-off aiming
> >> onto the gap between the raw, low level GPU/CUDA resources to bridge
> >> the very distance up to the Scikit-learn tools.
> >>
> >> Thought of a primitive mezzanine-approach for allowing minimum
> >> Scikit-learn integration efforts,
> >> i.e. to add the GPU-enabling layer into the Numpy, supposing all the
> >> Scikit-learn code uses vectorised Numpy services,
> >> to allow Numpy to call transparently GPU-services in case
> >> GPU-resources are detected on the <localhost>
> >> and to bypass the calls to CPU on such <localhost> that has no GPU
> >> resources available, which seems to be already available in iPython,
> >> if I remember well,
> >> but
> >> some GPU-grid approach is a principally better architecture for this.
> >>
> >> Appreciate your time, you´ve spent Andreas on reading this -- thank you.
> >>
> >> Remaining with respect,
> >> Michal Sourek
> >
> >
> >
> ------------------------------------------------------------------------------
> > Is your legacy SCM system holding you back? Join Perforce May 7 to find
> out:
> > • 3 signs your SCM is hindering your productivity
> > • Requirements for releasing software faster
> > • Expert tips and advice for migrating your SCM now
> > http://p.sf.net/sfu/perforce
> > _______________________________________________
> > Scikit-learn-general mailing list
> > Scikit-learn-general@lists.sourceforge.net
> > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
>
> ------------------------------------------------------------------------------
> Is your legacy SCM system holding you back? Join Perforce May 7 to find
> out:
> • 3 signs your SCM is hindering your productivity
> • Requirements for releasing software faster
> • Expert tips and advice for migrating your SCM now
> http://p.sf.net/sfu/perforce
> _______________________________________________
> Scikit-learn-general mailing list
> Scikit-learn-general@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
------------------------------------------------------------------------------
Is your legacy SCM system holding you back? Join Perforce May 7 to find out:
• 3 signs your SCM is hindering your productivity
• Requirements for releasing software faster
• Expert tips and advice for migrating your SCM now
http://p.sf.net/sfu/perforce
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general