def set_vb_param(self, phi_):
#unflatten and softmax
self.phi_ = phi_.reshape(self.N, self.K)
self.phi, logphi, self.H = softmax_weave(self.phi_)
self.phi_hat = self.phi.sum(0)
self.Hgrad = -logphi
if self.prior_Z == 'DP':
self.phi_tilde_plus_hat = self.phi_hat[::-1].cumsum()[::-1]
self.phi_tilde = self.phi_tilde_plus_hat - self.phi_hat
self.do_computations()
def set_vb_param(self,phi_):
#unflatten and softmax
self.phi_ = phi_.reshape(self.N,self.K)
self.phi, logphi, self.H = softmax_weave(self.phi_)
self.phi_hat = self.phi.sum(0)
self.Hgrad = -logphi
if self.prior_Z=='DP':
self.phi_tilde_plus_hat = self.phi_hat[::-1].cumsum()[::-1]
self.phi_tilde = self.phi_tilde_plus_hat - self.phi_hat
self.do_computations()
def set_vb_param(self,phi_):
"""
Accept a vector representing the variatinoal parameters, and reshape it into self.phi
"""
self.phi_ = phi_.reshape(self.N,self.K)
self.phi, logphi, self.H = softmax_weave(self.phi_)
self.phi_hat = self.phi.sum(0)
self.Hgrad = -logphi
if self.prior_Z=='DP':
self.phi_tilde_plus_hat = self.phi_hat[::-1].cumsum()[::-1]
self.phi_tilde = self.phi_tilde_plus_hat - self.phi_hat
self.do_computations()
def set_vb_param(self,phi_):
"""
Accept a vector representing the variatinoal parameters, and reshape it into self.phi
"""
self.phi_ = phi_.reshape(self.N, self.K)
self.phi, logphi, self.H = softmax_weave(self.phi_)
self.phi_hat = self.phi.sum(0)
self.Hgrad = -logphi
if self.prior_Z == 'DP':
self.phi_tilde_plus_hat = self.phi_hat[::-1].cumsum()[::-1]
self.phi_tilde = self.phi_tilde_plus_hat - self.phi_hat
self.do_computations()
def __init__(self, N, K, prior_Z='symmetric', alpha=1.0, name='col_mix'):
"""
Arguments
=========
N: the number of data
K: the (initial) number of cluster (or truncation)
prior_Z - either 'symmetric' or 'DP', specifies whether to use a symmetric Dirichlet prior for the clusters, or a (truncated) Dirichlet Process.
alpha: parameter of the Dirichelt (process)
"""
CollapsedVB.__init__(self, name)
self.N, self.K = N, K
assert prior_Z in ['symmetric','DP']
self.prior_Z = prior_Z
self.alpha = alpha
#random initial conditions for the vb parameters
self.phi_ = np.random.randn(self.N, self.K)
self.phi, logphi, self.H = softmax_weave(self.phi_)
self.phi_hat = self.phi.sum(0)
self.Hgrad = -logphi
if self.prior_Z == 'DP':
self.phi_tilde_plus_hat = self.phi_hat[::-1].cumsum()[::-1]
self.phi_tilde = self.phi_tilde_plus_hat - self.phi_hat
