sklearn_kernel.py

import sklearn.svm
import random
import numpy as np

if 0:
    import kernels
    kernel = kernels.histogram_intersection
else:
    kernel = lambda x, y: np.dot(x, y.T)


def sample_generator():
    num_points = 100
    num_dims = 10
    num_dims_half = num_dims / 2
    num_pos_points = num_points / 2

    x = np.random.random((num_points, num_dims))
    x[:num_pos_points, :num_dims_half] += .5
    x = (x.T / np.sum(x, 1)).T
    print(x[0])
    print(x[num_pos_points])
    y = np.zeros(num_points)
    y[:num_pos_points] = 1
    y = map(int, y)
    x = np.ascontiguousarray(x)
    y = np.ascontiguousarray(y)
    return x, y


x, y = sample_generator()
print(len(y))
print(x.shape)

svm = sklearn.svm.SVC(kernel=kernel)
svm.fit(x, y)


gram = kernel(x, x)
svm2 = sklearn.svm.SVC(kernel='precomputed')
svm2.fit(gram, y)

a, b = sample_generator()

dual_coef = svm.dual_coef_
intercept = svm.intercept_
support = svm.support_

for n, ax in zip(b, a):
    k = kernel(x, np.ascontiguousarray(ax.reshape((-1, 1)).T)).T
    dec_custom = np.sum(k[:, support] * dual_coef)
    d_custom = int(dec_custom.ravel()[0] < intercept[0])
    d = int(svm2.decision_function(k) >= svm2.intercept_)
    print('true_label[%d] predictions: svm.predict[%d] svm2.predict[%d] svm2.dec >= intercept[%d] manual[%d] decision_functions: svm.decision_function[%g] svm2.decision_function[%g] manual[%g]' % (n, svm.predict(ax)[0], svm2.predict(k)[0], d, d_custom, svm.decision_function(ax).ravel()[0], svm2.decision_function(k).ravel()[0], dec_custom.ravel()[0]))