def fit(self, X, Y):
n, d = X.shape[0], X.shape[1]
Y = Y.reshape(-1, 1)
L = self.calculateL(self.xi)
#secOrd = np.linalg.pinv((X.T.dot(X) * L + self.reg * np.eye(d)) / n)
# I = np.eye(d)
# I[-1, -1] = 0 #Not regularize the intercept
try:
secOrd = np.linalg.pinv(X.T.dot(X) * L / n + self.reg * np.eye(d))
except Exception as e:
print(e.message)
if not self.manualInitial:
self.initialize(X, Y)
t = 0.
while t < self.iterations:
v = X.dot(self._beta)
GEVFunc.clip(self.xi, v)
Y_hat = GEVFunc.inverseLink(self.xi, v)
#firOrd = (X.T.dot(Y_hat - Y) + self.reg * self._beta) / n
# tmp = self.reg * self._beta
# tmp[-1] = 0
firOrd = X.T.dot(Y_hat - Y) / n + self.reg * self._beta
deltaBeta = self.step * secOrd.dot(firOrd)
if t >= 100 and np.abs(deltaBeta).sum() < self.tol:
self._beta -= deltaBeta
# print "Converged. t = %d" % (t)
break
self._beta -= deltaBeta
t += 1
self._beta = np.array(self._beta).flatten()
def fit(self, X, Y):
self.__initialize(X, Y)
xi = self.xi
t = 0
while t < self.iterations:
#Caculate weight matrix
w = self._eta*np.power(-np.log(self._eta), xi+1)
W = np.diag(w)
#Z is used for updating beta
tmp = GEVFunc.derivLink(xi, self._eta)
Z = self._v + self._gamma \
*tmp \
*(Y - self._eta)
#Update beta
mat = np.matrix(X.T.dot(W).dot(X)) \
+ np.eye(X.shape[1])*self.regular
#self._beta = mat.I.dot(X.T).dot(W).dot(Z).getA1()
self._beta = np.linalg.pinv(mat).dot(X.T).dot(W).dot(Z).getA1()
#Calculate v
self._v = X.dot(self._beta)
GEVFunc.clip(xi, self._v)
#Judge if eta is convergent
newEta = GEVFunc.inverseLink2(xi, self._v)
if np.abs(newEta - self._eta).sum() < self.tol:
self._eta = newEta
break
self._eta = newEta
t += 1
def betaDeriva(self, X, Y, beta, xi):
v = X.dot(beta).reshape(-1, 1)
y = Y.reshape(-1, 1)
GEVFunc.clip(xi, v)
gev = GEVFunc.GEV(xi, v)
loggev = GEVFunc.logGEV(xi, v)
res = - np.sum(X * (loggev * (y - gev) / ((1 - gev) * (1 + xi*v))), axis=0) - beta
return res
def xiDeriva(self, X, Y, beta, xi):
v = X.dot(beta).reshape(-1, 1)
y = Y.reshape(-1, 1)
GEVFunc.clip(xi, v)
gev = GEVFunc.GEV(xi, v)
loggev = GEVFunc.logGEV(xi, v)
res = np.sum((np.log(1 + xi*v) / xi**2 - v / xi / (1 + xi*v))
* loggev * (y - gev) / (1 - gev)) - xi
return res
def oneStep3(self, X, Y):
n, d = X.shape[0], X.shape[1]
Y = Y.reshape(-1, 1)
v = X.dot(self._beta)
GEVFunc.clip(self.xi, v)
W = np.diag(GEVFunc.derivInverseLink(self.xi, v).flatten())
secOrd = np.linalg.pinv(X.T.dot(W).dot(X) / n + self.reg * np.eye(d))
Y_hat = GEVFunc.inverseLink(self.xi, v)
firOrd = X.T.dot(Y_hat - Y) / n + self.reg * self._beta
self._beta -= self.step * secOrd.dot(firOrd)
def betaSecDeriva(self, X, Y, beta, xi):
pos, neg = X[Y == 1], X[Y == 0]
v = pos.dot(beta).reshape(-1, 1)
GEVFunc.clip(xi, v)
w = (-1-xi) * np.power(1 + xi * v, -1./xi-2)
comp1 = pos.T.dot(np.diag(w.flatten())).dot(pos)
v = neg.dot(beta).reshape(-1, 1)
GEVFunc.clip(xi, v)
gev = GEVFunc.GEV(xi, v)
a = 1 + xi * v
w = np.power(a, -1./xi-2)*gev/(1-gev)*(1+xi - np.power(a, -1./xi)/(1-gev))
comp2 = neg.T.dot(np.diag(w.flatten())).dot(neg)
return comp1 + comp2 - np.eye(beta.size)
def oneStep(self, X, Y):
n, d = X.shape[0], X.shape[1]
Y = Y.reshape(-1, 1)
L = self.calculateL(self.xi)
try:
secOrd = np.linalg.pinv(X.T.dot(X) * L / n + self.reg * np.eye(d))
except Exception as e:
print(e.message)
v = X.dot(self._beta)
GEVFunc.clip(self.xi, v)
Y_hat = GEVFunc.inverseLink(self.xi, v)
firOrd = X.T.dot(Y_hat - Y) / n + self.reg * self._beta
self._beta -= self.step * secOrd.dot(firOrd)
def generateData(D=3, N=50, sigmaBeta=1, sigmaXi=1, miuXi=0.5):
flag = True
while flag:
X = np.random.randn(N, D)
beta = np.random.randn(D) * sigmaBeta
# xi = np.random.randn() * sigmaXi + miuXi
xi = miuXi
v = X.dot(beta)
GEVFunc.clip(xi, v)
y = GEVFunc.GEV(xi, v)
y[y < 0.5] = 0
y[y >= 0.5] = 1
if 0.3 < (y == 1).sum()/float(N) < 0.7:
flag = False
return X, y, xi, beta
def xiSecDeriva(self, X, Y, beta, xi):
v = X.dot(beta).reshape(-1, 1)
GEVFunc.clip(xi, v)
y = Y.reshape(-1, 1)
a = 1 + xi*v
gev = GEVFunc.GEV(xi, v)
loggev = GEVFunc.logGEV(xi, v)
y_gev = y - gev
one_gev = 1 - gev
y_gev_1_gev = y_gev/one_gev
xia = xi * a
xiv = xi * v
lna = np.log(a)
comp1 = (xiv * (3*a-1) - 2*lna*(a**2)) / (xi * xia**2) * y_gev_1_gev * loggev
comp2 = (1/xi**2 * lna - v/xia) * ((y-1)*loggev*gev/one_gev**2 + y_gev_1_gev) * (v/xia-lna/xi**2)/np.power(a, 1./xi)
return np.sum(comp1 + comp2) - 1
def fit3(self, X, Y):
n, d = X.shape[0], X.shape[1]
Y = Y.reshape(-1, 1)
if not self.manualInitial:
self.initialize(X, Y)
t = 0.
while t < self.iterations:
v = X.dot(self._beta)
GEVFunc.clip(self.xi, v)
W = np.diag(GEVFunc.derivInverseLink(self.xi, v).flatten())
# secOrd = np.linalg.pinv(X.T.dot(W).dot(X) / n + self.reg * np.eye(d))
A = X.T.dot(W).dot(X) / n + self.reg * np.eye(d)
Y_hat = GEVFunc.inverseLink(self.xi, v)
b = X.T.dot(Y_hat - Y) / n + self.reg * self._beta
self._beta += np.linalg.lstsq(A, -b)[0]
# self._beta -= self.step * secOrd.dot(firOrd)
t += 1
self._beta = np.array(self._beta).flatten()
def predict(self, X):
v = X.dot(self.beta)
GEVFunc.clip(self.xi, v)
return GEVFunc.inverseLink(self.xi, v).flatten()
def firstDeriva(self, X, Y, beta):
v = X.dot(beta)
GEVFunc.clip(self.xi, v)
Y_hat = GEVFunc.inverseLink(self.xi, v)
return X.T.dot(Y_hat - Y) / X.shape[0] + self.reg * beta
def predict(self, X):
self._v = X.dot(self._beta)
GEVFunc.clip(self.xi, self._v)
return GEVFunc.inverseLink2(self.xi, self._v)
def loglikelihood(self, beta, xi):
v = self.X.dot(beta.reshape(-1, 1))
GEVFunc.clip(xi, v)
res = GEVFunc.inverseLink(xi, v)
res[self.targets == 0] = 1 - res[self.targets == 0]
return sum([np.log(x + 1e-8) for x in res])