On 06/07/2013 09:33 PM, neo01124 wtf wrote:
Hi
I am using the scikitlearn implementation of Nu-SVR.
My problem (automatic phonetic segmentation for singing voice) has ~
50k points with 36 features. Seems relatively small to me compared to
the datasets I have been reading about. The problem is it takes a long
time (~6 hours) to fit the NuSVR model to 60% of this data and coarse
gridsearching for the parameters obviously takes even longer (a couple
of days). The times are with gridsearching with the njobs=-1 option
enabled on a 4 core machine.
How can I use joblibs to parallelise NuSVR ? Can it be done at all ?
Any other ideas to speedup what I am doing ?
The n_jobs=-1 means that the grid-search is in fact parallelized using
joblib. You should see that in your cpu usage.
Is there any particular reason to use nu-SVR instead of C-SVR? The C
version is a bit more standard, I think.
There also seems to be some issue: nu-SVR has a C parameter but not an
epsilon parameter. That doesn't make much sense to me?! I think we had a
discussion about this at some point but I'm confused again.
I am not very familiar with doing regression but I would expect the
runtime to be critically dependent on the kernel, C/nu and epsilon.
Training for hours using an rbf kernel on tens of thousands of points
seems the right order of magnitude for me.
You could probably increase epsilon / decrease C to make training faster.
Other than that, you could try a linear kernel first. That will be way
faster.
Hth,
Andy
------------------------------------------------------------------------------
How ServiceNow helps IT people transform IT departments:
1. A cloud service to automate IT design, transition and operations
2. Dashboards that offer high-level views of enterprise services
3. A single system of record for all IT processes
http://p.sf.net/sfu/servicenow-d2d-j
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
