2012/5/11 Gael Varoquaux <[email protected]>:
> On Fri, May 11, 2012 at 10:47:19PM +0900, Mathieu Blondel wrote:
>>    All algorithms which supports a warm_start constructor option should also
>>    be usable similarly to partial_fit. For example:
>
>>    from sklearn.linear_model import Lasso
>
>>    clf = Lasso(warm_start=True)
>>    clf.fit(X_subset1, y_subset1)
>>    clf.fit(X_subset2, y_subset2)
>>    ...
>
> I disagree: this is a very different thing than an online approach. In
> the example above, 'clf' reflects on the data of the second subset if the
> convergence has been successful.

+1 : the semantics of warm_start is *only to speedup the convergence*
by starting from a solution closer to the optimal solution of the
convex optimization problem (in this case the final solution will be
the solution of fit(X_subsetn, y_subsetn) ignoring the data from the
previous batches.

partial_fit has a different semantics: the data seen in the first
batches contribute to the final solution (generally using some kind of
online averaging sometimes using an explicit learning rate schedule).

-- 
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

------------------------------------------------------------------------------
Live Security Virtual Conference
Exclusive live event will cover all the ways today's security and 
threat landscape has changed and how IT managers can respond. Discussions 
will include endpoint security, mobile security and the latest in malware 
threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to