from __future__ import absolute_import
import numpy as np
from metric_learn.base_metric import BaseMetricLearner
def validate_cov_matrix(M):
M = (M + M.T) * 0.5
k = 0
I = np.eye(M.shape[0])
while True:
try:
_ = np.linalg.cholesky(M)
break
except np.linalg.LinAlgError:
# Find the nearest positive definite matrix for M. Modified from
# http://www.mathworks.com/matlabcentral/fileexchange/42885-nearestspd
# Might take several minutes
k += 1
w, v = np.linalg.eig(M)
min_eig = v.min()
M += (-min_eig * k * k + np.spacing(min_eig)) * I
return M
class KISSME(BaseMetricLearner):
def __init__(self):
self.M_ = None
def metric(self):
return self.M_
def fit(self, X, y=None):
n = X.shape[0]
if y is None:
y = np.arange(n)
X1, X2 = np.meshgrid(np.arange(n), np.arange(n))
X1, X2 = X1[X1 < X2], X2[X1 < X2]
matches = (y[X1] == y[X2])
num_matches = matches.sum()
num_non_matches = len(matches) - num_matches
idxa = X1[matches]
idxb = X2[matches]
S = X[idxa] - X[idxb]
C1 = S.transpose().dot(S) / num_matches
p = np.random.choice(num_non_matches, num_matches, replace=False)
idxa = X1[~matches]
idxb = X2[~matches]
idxa = idxa[p]
idxb = idxb[p]
S = X[idxa] - X[idxb]
C0 = S.transpose().dot(S) / num_matches
self.M_ = np.linalg.inv(C1) - np.linalg.inv(C0)
self.M_ = validate_cov_matrix(self.M_)
self.X_ = X
