def filterbank_matrices(self, a, mu_x, sigma2, epsilon=1e-9):
t_a, t_mu_x = align(a, mu_x)
temp = t_a - t_mu_x
temp, t_sigma = align(temp, sigma2)
temp = temp / (t_sigma * 2)
F = torch.exp(-torch.pow(temp, 2))
F = F / (F.sum(2, True).expand_as(F) + epsilon)
return F
def filterbank_matrices(self,a,mu_x,sigma2,epsilon=1e-9):
t_a,t_mu_x = align(a,mu_x)
temp = t_a - t_mu_x
temp,t_sigma = align(temp,sigma2)
temp = temp / (t_sigma * 2)
F = torch.exp(-torch.pow(temp,2))
# print 'F size:',F.size()
# F = F / (F.sum(2).expand_as(F) + epsilon)
Z=torch.unsqueeze(torch.unsqueeze(F.sum(2).sum(1),1),2).expand_as(F)
# print Z
F=F/Z
# print torch.sum(F)
# for batch in range(F.size()[0]):
# Z=torch.sum(F[batch,:,:])
# F[batch,:,:]=F[batch,:,:].clone()/Z
return F
def _filterbank_matrices(self, a, mu, sigma2, epsilon=1e-9):
"""Genrate filterbank matrix, call by function filterbank
Args:
a: x or y coordinates [0, self.A or self.B), size (1, 1, size.A or size.B)
mu_x: single-dimension mu of gaussian kernel, size (self.batch_size, self.N, 1)
sigma2: variance of gaussian kernel, size (self.batch_size, self.N, 1)
epsilon: minimun value to keep out 0-divisor
"""
t_a, t_mu = align(a, mu)
temp = t_a - t_mu
temp, t_sigma = align(temp, sigma2)
temp = temp / (t_sigma * 2)
F = torch.exp(-torch.pow(temp, 2))
# normalize for each kernel
F = F / (F.sum(2, True).expand_as(F) + epsilon)
return F