This aims at reproducing the sklearn.svm.SVC object without having to store 'support_vectors_' and '_dual_coef_'

svc_light_example.py

# Author: Jean-Remi King <[email protected]>
#
# License: BSD (3-clause)

import warnings
import numpy as np
import scipy.sparse as sp
from sklearn.svm import SVC, LinearSVC
from sklearn.datasets import make_classification
from sklearn.calibration import CalibratedClassifierCV

# CLASSIFIERS


def SVC_Light(probability=False, method='sigmoid', cv=5, **kwargs):
    """
    Similar to SVC(kernel='linear') without having to store 'support_vectors_'
     and '_dual_coef_'.
    Uses CalibrationClassifierCV if probability=True.
    """
    if probability is True:
        base_estimator = _SVC_Light(probability=True, **kwargs)
        return _SVC_Light_Proba(base_estimator=base_estimator, method=method,
                                cv=cv)
    else:
        return _SVC_Light(**kwargs)


class _SVC_Light_Proba(CalibratedClassifierCV):

    def decision_function(self, X):
        warnings.warn(
            "With 'probability=True' decision_function=predict_proba")
        return self.predict_proba(X)

    def fit(self, X, y):
        if len(np.unique(y)) > 2:
            # XXX
            raise ValueError('_SVC_Light currently does not support '
                             'probability=True for more than 2 classes.')
        super(_SVC_Light_Proba, self).fit(X, y)


class _SVC_Light(SVC):
    """
    Similar to SVC(kernel='linear') without having to store 'support_vectors_'
     and '_dual_coef_'
    """

    def __init__(self, kernel='linear', probability=False, **kwargs):
        if 'kernel' in kwargs.keys():
            raise ValueError('SVC_Light is only available when using a '
                             'linear kernel.')
        if 'probability' in kwargs.keys():
            raise RuntimeError('Currently, SVC_Light does not support '
                               'probability=True')
        super(_SVC_Light, self).__init__(kernel=kernel,
                                         probability=probability, **kwargs)

    def fit(self, X, y, scaling=None):
        super(_SVC_Light, self).fit(X, y)
        # compute coef from support vectors once only
        self._coef_ = self._compute_coef_()
        self.__delattr__('support_vectors_')
        self.__delattr__('_dual_coef_')

    def _compute_coef_(self):
        # Originally coef_(self) from SVC
        coef = self._get_coef()
        if sp.issparse(coef):
            coef.data.flags.writeable = False
        else:
            coef.flags.writeable = False
        return coef

    def predict(self, X):
        distances = self.decision_function(X)
        y_pred = predict_OneVsOne(distances, self.classes_)
        return y_pred

    def decision_function(self, X):
        X = self._validate_for_predict(X)
        n_sample = X.shape[0]
        intercept = np.tile(self.intercept_, (n_sample, 1))
        distances = np.dot(self.coef_, X.T).T + intercept
        if len(self.classes_) == 2:
            distances *= -1
        return distances

    @property
    def coef_(self):
        return self._coef_

# PREDICTERS


def predict_OneVsOne(confidence, classes):
    # for SVC, NuSVC
    n_samples, n_w = confidence.shape
    votes = np.zeros((n_samples, n_w))
    k = 0
    for i, class1 in enumerate(classes):
        for j, class2 in enumerate(classes[(i + 1):]):
            compared_classes = np.array([class1, class2])
            comparison = confidence[:, k] < 0
            votes[:, k] = compared_classes[comparison.astype(int)]
            k += 1
    summed_votes = np.array([np.sum(votes == c, axis=1) for c in classes]).T
    y_pred = predict_OneVsRest(summed_votes, classes)
    return y_pred


def predict_OneVsRest(confidence, classes):
    # for LinearSVC
    return np.array(classes[confidence.argmax(axis=1)])


# setup dataset --------------------------------------------------------------
X, y = make_classification(n_informative=10, n_classes=2)

# 1. Classic pipeline --------------------------------------------------------
svc = SVC(kernel='linear')
svc.fit(X, y)
y_pred = svc.decision_function(X)
score = svc.score(X, y)

# 2. Linear SVC ---------------------------------------------------------------
linearsvc = LinearSVC()
linearsvc.fit(X, y)
y_pred = linearsvc.decision_function(X)
score_linear = linearsvc.score(X, y)

# 3. Light SVC ----------------------------------------------------------------
svc_light = SVC_Light(kernel='linear')
svc_light.fit(X, y)
y_pred_light = svc_light.predict(X)
score_light = svc_light.score(X, y)

# 4. Light SVC Proba-----------------------------------------------------------
svc_light_proba = SVC_Light(kernel='linear', probability=True)
svc_light_proba.fit(X, y)
y_pred_light_proba = svc_light_proba.predict_proba(X)
y_pred_light_proba = svc_light_proba.predict(X)
score_light_proba = svc_light_proba.score(X, y)

print([score, score_linear, score_light, score_light_proba])