Ah, sorry, true. It is the error fraction instead of the number of errors.
In any case, try varying this quantity.
At one point I thought that nuSVC is just the constrained optimization
version of the lagrange-style (penalized) normal SVC. That would mean that
there is a correspondence between C for SVC and nu for nuSVC, leading to
the conclusion that there must be nus that are feasible. So setting to
nu=1. should always lead to feasibility. Now, looking at the docstring,
since the nu controls two quantities at the same time, I am not entirely
1000% sure of this anymore, but I think it still holds.
Michael
On Thu, Dec 8, 2016 at 11:08 AM, Piotr Bialecki <piotr.biale...@hotmail.de>
wrote:
> Hi Michael, hi Thomas,
>
> I think the nu value is bound to (0, 1].
> So the code will result in a ValueError (at least in sklearn 0.18).
>
> @Thomas
> I still think the optimization problem is not feasible due to your data.
> Have you tried balancing the dataset as I mentioned in your other question
> regarding the MLPClassifier?
>
>
> Greets,
> Piotr
>
>
>
>
>
>
> On 08.12.2016 10:57, Michael Eickenberg wrote:
>
> You have to set a bigger \nu.
> Try
>
> nus =2 ** np.arange(-1, 10) # starting at .5 (default), going to 512
> for nu in nus:
> clf = svm.NuSVC(nu=nu)
> try:
> clf.fit ...
> except ValueError as e:
> print("nu {} not feasible".format(nu))
>
> At some point it should start working.
>
> Hope that helps,
> Michael
>
>
>
>
> On Thu, Dec 8, 2016 at 10:49 AM, Thomas Evangelidis <teva...@gmail.com>
> wrote:
>
>> Hi Piotr,
>>
>> the SVC performs quite well, slightly better than random forests on the
>> same data. By training error do you mean this?
>>
>> clf = svm.SVC(probability=True)
>> clf.fit(train_list_resampled3, train_activity_list_resampled3)
>> print "training error=", clf.score(train_list_resampled3,
>> train_activity_list_resampled3)
>>
>> If this is what you mean by "skip the sample_weights":
>> clf = svm.NuSVC(probability=True)
>> clf.fit(train_list_resampled3, train_activity_list_resampled3,
>> sample_weight=None)
>>
>> then no, it does not converge. After all "sample_weight=None" is the
>> default value.
>>
>> I am out of ideas about what may be the problem.
>>
>> Thomas
>>
>>
>> On 8 December 2016 at 08:56, Piotr Bialecki < <piotr.biale...@hotmail.de>
>> piotr.biale...@hotmail.de> wrote:
>>
>>> Hi Thomas,
>>>
>>> the doc says, that nu gives an upper bound on the fraction of training
>>> errors and a lower bound of the fractions
>>> of support vectors.
>>> http://scikit-learn.org/stable/modules/generated/sklearn.svm.NuSVC.html
>>>
>>> Therefore, it acts as a hard bound on the allowed misclassification on
>>> your dataset.
>>>
>>> To me it seems as if the error bound is not feasible.
>>> How well did the SVC perform? What was your training error there?
>>>
>>> Will the NuSVC converge when you skip the sample_weights?
>>>
>>>
>>> Greets,
>>> Piotr
>>>
>>>
>>> On 08.12.2016 00:07, Thomas Evangelidis wrote:
>>>
>>> Greetings,
>>>
>>> I want to use the Nu-Support Vector Classifier with the following
>>> input data:
>>>
>>> X= [
>>> array([ 3.90387012, 1.60732281, -0.33315799, 4.02770896,
>>> 1.82337731, -0.74007214, 6.75989219, 3.68538903,
>>> ..................
>>> 0. , 11.64276776, 0. , 0. ]),
>>> array([ 3.36856769e+00, 1.48705816e+00, 4.28566992e-01,
>>> 3.35622071e+00, 1.64046508e+00, 5.66879661e-01,
>>> .....................
>>> 4.25335335e+00, 1.96508829e+00, 8.63453394e-06]),
>>> array([ 3.74986249e+00, 1.69060713e+00, -5.09921270e-01,
>>> 3.76320781e+00, 1.67664455e+00, -6.21126735e-01,
>>> ..........................
>>> 4.16700259e+00, 1.88688784e+00, 7.34729942e-06]),
>>> .......
>>> ]
>>>
>>> and
>>>
>>> Y= [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
>>> 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
>>> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
>>> 0, 0, 0, 0, ............................
>>> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
>>> 0, 0, 0, 0, 0, 0, 0, 0]
>>>
>>>
>>>> Each array of X contains 60 numbers and the dataset consists of 48
>>>> positive and 1230 negative observations. When I train an svm.SVC()
>>>> classifier I get quite good predictions, but wit the svm.NuSVC() I keep
>>>> getting the following error no matter which value of nu in [0.1, ..., 0.9,
>>>> 0.99, 0.999, 0.9999] I try:
>>>> /usr/local/lib/python2.7/dist-packages/sklearn/svm/base.pyc in
>>>> fit(self, X, y, sample_weight)
>>>> 187
>>>> 188 seed = rnd.randint(np.iinfo('i').max)
>>>> --> 189 fit(X, y, sample_weight, solver_type, kernel,
>>>> random_seed=seed)
>>>> 190 # see comment on the other call to np.iinfo in this file
>>>> 191
>>>> /usr/local/lib/python2.7/dist-packages/sklearn/svm/base.pyc in
>>>> _dense_fit(self, X, y, sample_weight, solver_type, kernel, random_seed)
>>>> 254 cache_size=self.cache_size, coef0=self.coef0,
>>>> 255 gamma=self._gamma, epsilon=self.epsilon,
>>>> --> 256 max_iter=self.max_iter, random_seed=random_seed)
>>>> 257
>>>> 258 self._warn_from_fit_status()
>>>> /usr/local/lib/python2.7/dist-packages/sklearn/svm/libsvm.so in
>>>> sklearn.svm.libsvm.fit (sklearn/svm/libsvm.c:2501)()
>>>> ValueError: specified nu is infeasible
>>>
>>>
>>>
>>> Does anyone know what might be wrong? Could it be the input data?
>>>
>>> thanks in advance for any advice
>>> Thomas
>>>
>>>
>>>
>>> --
>>>
>>> ======================================================================
>>>
>>> Thomas Evangelidis
>>>
>>> Research Specialist
>>> CEITEC - Central European Institute of Technology
>>> Masaryk University
>>> Kamenice 5/A35/1S081,
>>> 62500 Brno, Czech Republic
>>>
>>> email: tev...@pharm.uoa.gr
>>>
>>> teva...@gmail.com
>>>
>>>
>>> website: <https://sites.google.com/site/thomasevangelidishomepage/>
>>> https://sites.google.com/site/thomasevangelidishomepage/
>>>
>>>
>>>
>>> _______________________________________________
>>> scikit-learn mailing
>>> listscikit-learn@python.orghttps://mail.python.org/mailman/listinfo/scikit-learn
>>>
>>> _______________________________________________ scikit-learn mailing
>>> list scikit-learn@python.org https://mail.python.org/mailma
>>> n/listinfo/scikit-learn
>>
>> --
>>
>> ======================================================================
>>
>> Thomas Evangelidis
>>
>> Research Specialist
>> CEITEC - Central European Institute of Technology Masaryk University
>> Kamenice 5/A35/1S081, 62500 Brno, Czech Republic
>>
>> email: tev...@pharm.uoa.gr
>>
>> teva...@gmail.com
>>
>> website: https://sites.google.com/site/thomasevangelidishomepage/
>>
>> _______________________________________________ scikit-learn mailing
>> list scikit-learn@python.org https://mail.python.org/mailma
>> n/listinfo/scikit-learn
>
> _______________________________________________
> scikit-learn mailing
> listscikit-learn@python.orghttps://mail.python.org/mailman/listinfo/scikit-learn
>
>
> _______________________________________________
> scikit-learn mailing list
> scikit-learn@python.org
> https://mail.python.org/mailman/listinfo/scikit-learn
>
>
_______________________________________________
scikit-learn mailing list
scikit-learn@python.org
https://mail.python.org/mailman/listinfo/scikit-learn
