Cross validation for custom kernel SVM in scikit-learn

Question

I would like to do a grid-search through cross-validation for a custom kernel SVM using scikit-learn. More precisely following this example I want to define a kernel function like

def my_kernel(x, y):
"""
We create a custom kernel:
k(x, y) = x * M *y.T          
"""
return np.dot(np.dot(x, M), y.T)

where M is a parameter of the kernel (like the gamma in the gaussian kernel).

I want to feed this parameter M through GridSearchCV, with something like

parameters = {'kernel':('my_kernel'), 'C':[1, 10], 'M':[M1,M2]}
svr = svm.SVC()
clf = grid_search.GridSearchCV(svr, parameters)

So my question is : how to define the my_kernel so that the M variable will be given by GridSearchCV ?

Andreus Andreus · Accepted Answer · 2015-07-13T13:49:05

You may have to make a wrapper class. Something like:

class MySVC(BaseEstimator,ClassifierMixin):
    def __init__( self, 
              # all the SVC attributes
              M ):
         self.M = M
         # etc...

    def fit( self, X, y ):
         kernel = lambda x,y : np.dot(np.dot(x,M),y.T)
         self.svc_ = SVC( kernel=kernel, # the other parameters )
         return self.svc_.fit( X, y )
    def predict( self, X ):
         return self.svc_.predict( X )
    # et cetera

Cross validation for custom kernel SVM in scikit-learn

1 Answers