Hi all,
I know that there are some other people out there using scikit-learn for
MVPA with neuroimaging data. Does anyone have an implementation of the
searchlight method (i.e. http://www.ncbi.nlm.nih.gov/pubmed/16537458) using
scikit-learn that they would be willing to share?
Michael
--
Good to see that we reached a consensus. So, predict_scores will be a
wrapper around existing score-prediction methods and we'll be able to
use it in other places such KMeans. I won't have time to implement
this proposal any time soon. If anyone feels like doing it, feel free
to jump in.
Mathieu
On Thu, Sep 29, 2011 at 05:59:49PM +0200, Olivier Grisel wrote:
> Ok for predict_score then.
predict_scores, with an 's' at the end.
G
--
All the data continuously generated in your IT infrastructure contains a
definitiv
Ok for predict_score then.
--
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel
--
All the data continuously generated in your IT infrastructure contains a
definitive record of customers, application perform
+1
-Original Message-
From: Paolo Losi
Date: Thu, 29 Sep 2011 17:57:17
To:
Reply-To: scikit-learn-general@lists.sourceforge.net
Subject: Re: [Scikit-learn-general] Unifying predict_proba,
predict_log_proba and decision_function
--
On Thu, Sep 29, 2011 at 5:54 PM, Gael Varoquaux <
gael.varoqu...@normalesup.org> wrote:
> On Thu, Sep 29, 2011 at 11:53:20AM -0400, Alexandre Gramfort wrote:
> > predict_scores?
>
> ^ my favorite
+1
--
All the data conti
On Thu, Sep 29, 2011 at 11:53:20AM -0400, Alexandre Gramfort wrote:
> predict_scores?
^ my favorite.
G
--
All the data continuously generated in your IT infrastructure contains a
definitive record of customers, applicati
predict_scores?
predict_sample_scores?
predict_values?
Alex
On Thu, Sep 29, 2011 at 11:47 AM, Mathieu Blondel wrote:
> On Fri, Sep 30, 2011 at 12:38 AM, Olivier Grisel
> wrote:
>
>> "score" seems a bit too generic to me and I think it would conflict
>> with our existing use of the word "score"
Hi Olivier,
I would favour score with respect to confidence.
Score is frequently used in probability calibration papers.
See for example:
"Transforming classifier scores into accurate multiclass probability
estimates"
by Zadrozny
Paolo
On Thu, Sep 29, 2011 at 5:38 PM, Olivier Grisel wrote:
>
On Fri, Sep 30, 2011 at 12:38 AM, Olivier Grisel
wrote:
> "score" seems a bit too generic to me and I think it would conflict
> with our existing use of the word "score" in the metrics module which
> is more about the general goodness of fit of an estimated model w.r.t.
> a complete datasets (e.g
> +1 for the general idea, but what about using "predict_confidence"
> instead of "predict_score"?
-1 as I feel predict_confidence has a specific meaning in stats
and score makes things clear (the bigger the better)
My concern is the same naming for estimator.score(X) and
estimator.predict_score(
On Thu, Sep 29, 2011 at 11:38:09AM -0400, Alexandre Gramfort wrote:
> > I hadn't thought about it but it's a good idea. We could drop the
> > leading underscore (_predict_margin -> predict_margin), this way we
> > can keep using hasattr(clf, "predict_margin") to check the
> > capabilities of an obj
+1 for the general idea, but what about using "predict_confidence"
instead of "predict_score"?
As far as I can tell "confidence" is often the case that confidence
can be some sort of loosely defined unormalized probability. But maybe
it conflicts with the meaning of "confidence" as in "confidence
> I hadn't thought about it but it's a good idea. We could drop the
> leading underscore (_predict_margin -> predict_margin), this way we
> can keep using hasattr(clf, "predict_margin") to check the
> capabilities of an object (Gael really likes this and it's true that
> it's nice!). That would als
On Fri, Sep 30, 2011 at 12:32:42AM +0900, Mathieu Blondel wrote:
> On Fri, Sep 30, 2011 at 12:16 AM, Lars Buitinck wrote:
> > def predict_score(self, X, score):
> > if score == 'margin':
> > return self._predict_margin(X)
> > etc.?
> I hadn't thought about it but it's a good idea. We
On Thu, Sep 29, 2011 at 03:48:22PM +0200, Olivier Grisel wrote:
> 2011/9/29 Lars Buitinck :
> > 2011/9/29 Peter Prettenhofer :
> >> +1, we should change the behavior of SGDClassifier to match SVC
> > I see now that SVC(probability=True) indeed returns (n_samples,
> > n_classes) output... any othe
On Thu, Sep 29, 2011 at 03:47:09PM +0200, Olivier Grisel wrote:
> > LinearSVC.predict_proba raises a NotImplementedError ...
> Yes and I think this is a bad pattern. We should remove the
> predict_proba method from LinearSVC. We should remove it from the
> liblinear base class and just define it f
On Fri, Sep 30, 2011 at 12:16 AM, Lars Buitinck wrote:
> def predict_score(self, X, score):
> if score == 'margin':
> return self._predict_margin(X)
>
> etc.?
I hadn't thought about it but it's a good idea. We could drop the
leading underscore (_predict_margin -> predict_margin), this
2011/9/29 Mathieu Blondel :
> LogisticRegression supports all three kinds of score (proba, log proba
> and margin), therefore we could have clf.predict_score(X,
> score="margin|proba|log_proba").
I like the idea in principle. I assume we would implement this in a
base class or mixin as something l
+10
1) this solution standardizes the handling of the score concept
as Mathieu suggested.
2) makes implementing score calibration to get probability estimates
indipendent from the classifier.
Paolo
On Thu, Sep 29, 2011 at 4:54 PM, Mathieu Blondel wrote:
> Ok, I'm starting a new thread t
sounds good but I'd rather have shape [n_samples, n_classes]
so the indexing axis is the same as for X.
Alex
On Thu, Sep 29, 2011 at 10:54 AM, Mathieu Blondel wrote:
> Ok, I'm starting a new thread to throw a new idea.
>
> Several times I needed to implement a method that could generalize to
>
Ok, I'm starting a new thread to throw a new idea.
Several times I needed to implement a method that could generalize to
new inputs and predict a score, which is however not a probability.
decision_function is an example of such a method but the name is not
really good and too much tight to linear
>> does decision_function make sense in the multiclass case?
>
> It does make sense with one-vs-all (LinearSVC). One-vs-one (SVC) is
> more problematic...
ok it suggests that decision_function is an ill-posed concept and
that's it's estimator dependent.
>> maybe with an np.argmax(df, axis=1) ?
>
does decision_function make sense in the multiclass case?
maybe with an np.argmax(df, axis=1) ?
in that case I think the output should be (n_samples, 1)
if not (n_samples,)
Alex
On Thu, Sep 29, 2011 at 9:53 AM, Mathieu Blondel wrote:
> On Thu, Sep 29, 2011 at 10:48 PM, Olivier Grisel
> wrote:
2011/9/29 Mathieu Blondel :
> On Thu, Sep 29, 2011 at 10:48 PM, Olivier Grisel
> wrote:
>
>> +1 for the (n_samples, n_classes) shape with probabilities that sum to
>> 1.0 for all predict_proba implementations (binary or multiclass).
>
> In that case, what about decision_function? Currently, in th
On Thu, Sep 29, 2011 at 10:48 PM, Olivier Grisel
wrote:
> +1 for the (n_samples, n_classes) shape with probabilities that sum to
> 1.0 for all predict_proba implementations (binary or multiclass).
In that case, what about decision_function? Currently, in the binary
case, it returns (1, n_sample
2011/9/29 Olivier Grisel :
> 2011/9/29 Lars Buitinck :
>> 2011/9/29 Peter Prettenhofer :
>>>
>>> +1, we should change the behavior of SGDClassifier to match SVC
>>
>> I see now that SVC(probability=True) indeed returns (n_samples,
>> n_classes) output... any other opinions on this?
>
> +1 for the (
could you be more explicit about the change of behavior?
does it match with the LibLinear predict_proba?
>>>
>>> LinearSVC doesn't have predict_proba,
>>
>> it does for the logistic regression.
>
> LinearSVC.predict_proba raises a NotImplementedError ...
LibLinear if also used for t
2011/9/29 Lars Buitinck :
> 2011/9/29 Peter Prettenhofer :
>>
>> +1, we should change the behavior of SGDClassifier to match SVC
>
> I see now that SVC(probability=True) indeed returns (n_samples,
> n_classes) output... any other opinions on this?
+1 for the (n_samples, n_classes) shape with proba
2011/9/29 Peter Prettenhofer :
> 2011/9/29 Alexandre Gramfort :
could you be more explicit about the change of behavior?
does it match with the LibLinear predict_proba?
>>>
>>> LinearSVC doesn't have predict_proba,
>>
>> it does for the logistic regression.
>
> LinearSVC.predict_prob
2011/9/29 Peter Prettenhofer :
>
> +1, we should change the behavior of SGDClassifier to match SVC
I see now that SVC(probability=True) indeed returns (n_samples,
n_classes) output... any other opinions on this?
--
Lars Buitinck
Scientific programmer, ILPS
University of Amsterdam
--
2011/9/29 Alexandre Gramfort :
>>> could you be more explicit about the change of behavior?
>>>
>>> does it match with the LibLinear predict_proba?
>>
>> LinearSVC doesn't have predict_proba,
>
> it does for the logistic regression.
LinearSVC.predict_proba raises a NotImplementedError ...
>
>> bu
>> could you be more explicit about the change of behavior?
>>
>> does it match with the LibLinear predict_proba?
>
> LinearSVC doesn't have predict_proba,
it does for the logistic regression.
> but it matches the
> SGDClassifier's: 1-d output containing the probabilities of the
> positive class
2011/9/29 Alexandre Gramfort :
> could you be more explicit about the change of behavior?
>
> does it match with the LibLinear predict_proba?
LinearSVC doesn't have predict_proba, but it matches the
SGDClassifier's: 1-d output containing the probabilities of the
positive class in the binary case.
could you be more explicit about the change of behavior?
does it match with the LibLinear predict_proba?
Alex
On Thu, Sep 29, 2011 at 9:21 AM, Mathieu Blondel wrote:
> On Thu, Sep 29, 2011 at 10:16 PM, Lars Buitinck wrote:
>
>> Mathieu's multiclass.predict_binary relied on the old
>> Multinomi
On Thu, Sep 29, 2011 at 10:16 PM, Lars Buitinck wrote:
> Mathieu's multiclass.predict_binary relied on the old
> MultinomialNB.predict_proba behavior. Fixed it.
Thanks. Do all classifiers with a predict_proba method comply to the
new behavior?
Mathieu
--
2011/9/29 Olivier Grisel :
> Hi all,
>
> builbot keeps sending me emails about broken tests in master:
>
> http://buildbot.afpy.org/scikit-learn/builders/ubuntu-64bit/builds/8340/steps/test/logs/stdio
Mathieu's multiclass.predict_binary relied on the old
MultinomialNB.predict_proba behavior. Fixed
Hi all,
builbot keeps sending me emails about broken tests in master:
http://buildbot.afpy.org/scikit-learn/builders/ubuntu-64bit/builds/8340/steps/test/logs/stdio
==
ERROR: sklearn.tests.test_multiclass.test_ovr_fit_predict
---
38 matches
Mail list logo
