[scikit-learn] NuSVC and ValueError: specified nu is infeasible

Michael Eickenberg michael.eickenberg at gmail.com
Thu Dec 8 05:18:17 EST 2016 

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.bialecki at 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 <tevang3 at 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.bialecki at hotmail.de>
>> piotr.bialecki at 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: tevang at pharm.uoa.gr
>>>
>>>           tevang3 at gmail.com
>>>
>>>
>>> website: <https://sites.google.com/site/thomasevangelidishomepage/>
>>> https://sites.google.com/site/thomasevangelidishomepage/
>>>
>>>
>>>
>>> _______________________________________________
>>> scikit-learn mailing listscikit-learn at python.orghttps://mail.python.org/mailman/listinfo/scikit-learn
>>>
>>> _______________________________________________ scikit-learn mailing
>>> list scikit-learn at 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: tevang at pharm.uoa.gr
>>
>>           tevang3 at gmail.com
>>
>> website: https://sites.google.com/site/thomasevangelidishomepage/
>>
>> _______________________________________________ scikit-learn mailing
>> list scikit-learn at python.org https://mail.python.org/mailma
>> n/listinfo/scikit-learn
>
> _______________________________________________
> scikit-learn mailing listscikit-learn at python.orghttps://mail.python.org/mailman/listinfo/scikit-learn
>
>
> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-learn/attachments/20161208/8097c8c4/attachment-0001.html>

More information about the scikit-learn mailing list