#!/usr/bin/env python

from sklearn.svm import SVC
import numpy as np

#Set up toy SVC problem
X = np.random.random(size=(1000,2))
Y = [1.0 if x1 + x2 > 1 else -1.0 for x1, x2 in X]
C = 10e9

#Calculate the constant term for linear SVC using alphas, support vectors, and weights
clf = SVC(C=C, kernel='linear', verbose=True)
clf.fit(X, Y)

nSV = sum(clf.n_support_)
b, soft = 0.0, 0.0
for m in range(nSV):
  if np.abs(clf.dual_coef_[0][m]) < C:
    b += (Y[clf.support_[m]] - np.dot(X[clf.support_[m]], np.transpose(clf.coef_)))
  else:
    soft += 1
b /= (nSV - soft)
print "b = %f, clf.intercept_ = %f" % (b, clf.intercept_)

#Calculate the constant term for polynomial SVC using alphas and support vectors
coef0, Q = 1.0, 3.0
Kpoly = lambda X, X_dash: (np.dot(X, X_dash) + coef0)**Q
clf = SVC(C=C, kernel='poly', coef0=coef0, degree=Q, verbose=True)
clf.fit(X, Y)

nSV = sum(clf.n_support_)
b, soft = 0.0, 0.0
for m in range(nSV):
  if np.abs(clf.dual_coef_[0][m]) < C:
    b += (Y[clf.support_[m]] - np.dot(clf.dual_coef_,
                    [Kpoly(sv, X[clf.support_[m]]) for sv in clf.support_vectors_]))
  else:
    soft += 1
b /= (nSV - soft)
print "b = %f, clf.intercept_ = %f" % (b, clf.intercept_)