def MSE(self, X, Y, delimiter=" "):
"""Mean Squared Error (or mis-classification error).
"""
MSE = 0
genX, _, N = batchX(X, self.batch, delimiter)
genT, _ = batchT(Y, self.batch, delimiter, self.C_dict)
for X, T in zip(genX, genT):
H = np.dot(X, self.W)
for i in xrange(H.shape[1]):
H[:, i] = self.ufunc[i](H[:, i])
Th1 = H.dot(self.B)
p = float(X.shape[0]) / N
MSE += mse(T, Th1, self.classification, self.multiclass) * p
return MSE
def train(self, X, T, delimiter=" "):
"""Trains ELM, can use any X and T(=Y), and specify neurons.
Neurons: (number, type, [W], [B])
"""
# get parameters of new data and add neurons
self.Xmean, self.Xstd = meanstdX(X, self.batch, delimiter)
if self.classification:
self.C_dict = c_dictT(T, self.batch)
# get data iterators
genX, self.inputs, N = batchX(X, self.batch, delimiter)
genT, self.targets = batchT(T, self.batch, delimiter, self.C_dict)
# get mean value of targets
if self.classification or self.multiclass:
self.Tmean = np.zeros((self.targets,)) # for any classification
else:
self.Tmean, _ = meanstdX(T, self.batch, delimiter)
# project data
nn = len(self.ufunc)
HH = np.zeros((nn, nn))
HT = np.zeros((nn, self.targets))
for X, T in zip(genX, genT):
# get hidden layer outputs
H = np.dot(X, self.W)
for i in xrange(H.shape[1]):
H[:, i] = self.ufunc[i](H[:, i])
H, T = semi_Tikhonov(H, T, self.Tmean) # add Tikhonov regularization
# least squares solution - multiply both sides by H'
p = float(X.shape[0]) / N
HH += np.dot(H.T, H)*p
HT += np.dot(H.T, T)*p
# solve ELM model
HH += self.cI * np.eye(nn) # enhance solution stability
self.B = lstsq(HH, HT)[0]
