Use StratifiedKFold
On 12 September 2014 13:03, Pagliari, Roberto
wrote:
> When using SVM or linearSVC, is it possible to force
> cross_validation.KFold to generate subsets with both classes (in the case
> of a two-class problem)?
>
>
> --
When using SVM or linearSVC, is it possible to force cross_validation.KFold to
generate subsets with both classes (in the case of a two-class problem)?
--
Want excitement?
Manually upgrade your production database.
When yo
Good point. It should be straightforward in any case, something like:
class Quantizer(sklearn.base.TransformerMixin):
def __init__(self, thresholds):
self.thresholds = thresholds # must be sorted
def transform(X, y=None):
return np.searchsorted(self.thresholds, X)
On 12 S
In my case I would like to do it right after scaling, while doing grid search.
This would be different to quantize the entire training set at the beginning.
Thank you,
From: Joel Nothman [mailto:[email protected]]
Sent: Thursday, September 11, 2014 9:00 PM
To: scikit-learn-general
Subject
If thresholds can be provided to the constructor then they are not
estimated automatically from the training data. This is the sort of
preprocessing you can and should do with pandas.
On 12 September 2014 10:53, Pagliari, Roberto
wrote:
> I’m getting errors about get_params_ missing etc…
>
>
>
I’m getting errors about get_params_ missing etc…
I guess I need to derive my own binarizer from some other classes. Is there a
way to simplify the process?
Essentially, what I need is the binarizer, with more levels (and thresholds
provided to the constructors).
Thank you
From: Joel Nothman
For quantizing or binning? Not currently.
On 12 September 2014 06:31, Pagliari, Roberto
wrote:
> Is there something like the binarizer with more levels (thresholds
> provided with input)
>
>
>
>
>
> Thanks
>
>
>
>
> --
That appears to work (with a small modification):
param_grid = [{'eta0':[1/alpha_this_step], 'alpha':[alpha_this_step]} for
alpha_this_step in 10.0**(np.arange(11)-5)]
Neat trick. Thanks.
Unfortunately I don't have time to do the experiment, but eta = 1/t and eta
= 1/(alpha*t) seem to be pretty
Is there something like the binarizer with more levels (thresholds provided
with input)
Thanks
--
Want excitement?
Manually upgrade your production database.
When you want reliability, choose Perforce
Perforce version c
You can actually do that using the current grid-search.
Specify the "grid" as a list of single grid-points. That should do.
param_grid = [{'eta0':1/alpha_this_step, 'alpha':alpha_this_step} for
alpha_this_step in my_alphas]
That should do it, right?
I think for the "optimum" the same guarantee
My mistake. I was using those names in the wrong place.
Thank you,
-Original Message-
From: Andy [mailto:[email protected]]
Sent: Thursday, September 11, 2014 4:24 PM
To: [email protected]
Subject: Re: [Scikit-learn-general] modify gridsearch to scale cross-valid
On 09/11/2014 07:58 PM, Pagliari, Roberto wrote:
> I'm getting errors when using these parameters
>
> linear_svm__penalty
> linear_svm__loss
> linear_svm__dual
>
> don't they have the same names?
They do.
Can you provide a snipplet and a traceback?
-
Just noticed I got my Greeks wrong. By nu I meant eta everywhere.
On Thu, Sep 11, 2014 at 2:24 PM, F. William High wrote:
> The Shalev-Schwartz et al. Pegasos update rule on the learning rate
> parameter is
>
> nu_i = 1 / (lambda * i)
>
> where lambda multiplies the regularization term. If thi
This is simply due to the fact that centering a sparse matrix (that is
subtracting from its mean), would make it dense, and there would be no
point of it being sparse in the first place.
On Thu, Sep 11, 2014 at 9:09 PM, Pagliari, Roberto
wrote:
> This is what I’m getting when using sparse matric
This is what I'm getting when using sparse matrices with grid search:
Cannot center sparse matrices: pass `with_mean=False` instead. See docstring
for motivation and alternatives.
If I set with_mean=False, does that mean the data will not be scaled with
respect to the mean? If so, why would one
The Shalev-Schwartz et al. Pegasos update rule on the learning rate
parameter is
nu_i = 1 / (lambda * i)
where lambda multiplies the regularization term. If this rule is used,
they show you can converge to an error of epsilon in O(1/(lambda*epsilon))
iterations at high probability. This differs
You'll want a predict() or a transform() as well, depending on whether it's
acting as an estimator or transformer.
On Thu, Sep 11, 2014 at 9:57 AM, Pagliari, Roberto
wrote:
> Thank you,
>
> I’m going to try shortly.
>
>
>
> And in general, if I wanted to put my own function in the pipeline, the
I'm getting errors when using these parameters
linear_svm__penalty
linear_svm__loss
linear_svm__dual
don't they have the same names?
I tried linear_svc but it doesn't work either
Thank you
-Original Message-
From: Laurent Direr [mailto:[email protected]]
Sent: Thursday,
Thank you,
I’m going to try shortly.
And in general, if I wanted to put my own function in the pipeline, the only
requirement is that the class must have the “fit” method?
Thank you again,
From: Josh Vredevoogd [mailto:[email protected]]
Sent: Thursday, September 11, 2014 12:52 PM
To: sciki
Hi,
If you test this code you will see it raises an error ;).
The naming of the parameters in the param_grid should be consistent
with the names in the Pipeline object.
GridSearchCV performs grid search on the Pipeline object so it cannot
understand what the 'LinearSVC__C' parameter means.
If
You're missing estimators = in the first line, I guess.
params should be:
params = dict(linear_svm__C=[0.1, 10, 100])
On Thu, Sep 11, 2014 at 9:44 AM, Pagliari, Roberto
wrote:
> Hi,
> Yes, I think you are right.
>
> Is the code below how it should be done (scaling+linearsvc)?
>
> [('scaler', Sca
Hi,
Yes, I think you are right.
Is the code below how it should be done (scaling+linearsvc)?
[('scaler', Scaler()), ('linear_svm', LinearSVC())]
clf = Pipeline(estimators)
params = dict(LinearSVC__C=[0.1, 10, 100])
gs = GridSearchCV(clf, param_grid=params)
Thank you,
-Original Message---
Hello,
I think a pipeline does precisely what you are asking for:
http://scikit-learn.org/stable/modules/pipeline.html
If you include the scaler as a step in the pipeline it should behave the
way you described in your first email.
Laurent
On 09/11/2014 04:59 PM, Pagliari, Roberto wrote:
> I'm
I'm not trying to scale the dataset at the very beginning. I would like to
scale while doing gridsearchCV.
Thanks,
-Original Message-
From: Pagliari, Roberto [mailto:[email protected]]
Sent: Thursday, September 11, 2014 10:52 AM
To: [email protected]
Subj
I'm not sure how to do it when using gridsearch. Can you provide an example?
Thank you,
-Original Message-
From: Gael Varoquaux [mailto:[email protected]]
Sent: Thursday, September 11, 2014 10:50 AM
To: [email protected]
Subject: Re: [Scikit-learn-ge
Use a pipeline.
G
On Thu, Sep 11, 2014 at 02:47:48PM +, Pagliari, Roberto wrote:
> Hello,
> Gridsearch with CV is something like this at a high level:
> for every combination of parameters:
>for every partition of training data
> split training into train_cv and test_cv
>
Hello,
Gridsearch with CV is something like this at a high level:
for every combination of parameters:
for every partition of training data
split training into train_cv and test_cv
train_classifier(train_cv).predict(test_cv)
compute score
average score
if max so far, then u
27 matches
Mail list logo
