TL;DNR: parameter search results datastructure choice should
anticipate new use-cases
Thanks Joel for the detailed analysis.
I the current situation I think I my-self I like:
5. many attributes, each an array, on a custom results object
This makes it possible to write a `__repr__` method on tha
It would be great if there was a way to access the parameter search results as
a numpy ndarray, with one axis for each parameter and one additional axis for
the cross-validation folds. This would make it easy to visualise the grid
search results, compute the mean, median or variance for each gri
On 06/07/2013 03:13 PM, Romaniuk, Michal wrote:
> It would be great if there was a way to access the parameter search results
> as a numpy ndarray, with one axis for each parameter and one additional axis
> for the cross-validation folds. This would make it easy to visualise the grid
> search re
On 06/07/2013 12:08 AM, Joel Nothman wrote:
> I proposed something that did this among a more general solution for
> warm starts without memoizing a couple of weeks ago, but I think
> memoizing is neater and handles most cases. To handle it generally,
> you could add a memoize parameter to Pipel
> Memorization and parallelization don't play along nicely.
Yes, I am strongly thinking of adding optional memoization directly to
joblib.Parallel. It is often a fairly natural place to put a memoization
as structures should be pickleable and data transfer should be limited.
What do people think?
> It would be great if there was a way to access the parameter search
> results as a numpy ndarray, with one axis for each parameter and one
> additional axis for the cross-validation folds. This would make it easy
> to visualise the grid search results, compute the mean, median or
> variance for e
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 NuSV
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 i
**Kernel SVM are not scalable** to large or even medium number of
samples as the complexity is quadratic (or more). You should try to:
- learn independent SVR models on a partitions of the data (e.g. 10
models trained on 5000 samples each) and then compute the mean
predictions of the 10 models as
On May 23, 2013, at 5:03 AM, Gilles Louppe wrote:
>> So I'd like to contribute a simple MAE criterion that would be efficient for
>> random splits (i.e. O(n) given a single batch update.) Is the direction
>> forward for something like this to hard-code more criteria in _tree.pyx, or
>> would it
10 matches
Mail list logo
