class MFMixGaussian(Likelihood):
"""
q(z | lambda ) = q(pi) prod_{k=1}^K q(mu_k) q(sigma_k)
q(pi) = Dirichlet(alpha')
q(mu_k) = N(mu'_k, Sigma'_k)
q(sigma_k) = Inv-Gamma(a'_k, b'_k)
where z = {pi, mu, sigma}
"""
def __init__(self, K, D):
self.K = K
self.D = D
self.dirichlet = MFDirichlet(1, K)
self.gaussian = MFGaussian(K*D)
self.invgamma = MFInvGamma(K*D)
Likelihood.__init__(self, self.dirichlet.num_vars + \
self.gaussian.num_vars + \
self.invgamma.num_vars)
self.num_params = self.dirichlet.num_params + \
self.gaussian.num_params + \
self.invgamma.num_params
def mapping(self, x):
return [self.dirichlet.mapping(x),
self.gaussian.mapping(x),
self.invgamma.mapping(x)]
def set_params(self, params):
self.dirichlet.set_params(params[0])
self.gaussian.set_params(params[1])
self.invgamma.set_params(params[2])
def print_params(self, sess):
self.dirichlet.print_params(sess)
self.gaussian.print_params(sess)
self.invgamma.print_params(sess)
def sample(self, size, sess):
"""z ~ q(z | lambda)"""
z_dirichlet = self.dirichlet.sample((size[0], self.dirichlet.num_vars), sess)
z_gaussian = sess.run(self.gaussian.sample((size[0], self.gaussian.num_vars), sess))
z_invgamma = self.invgamma.sample((size[0], self.invgamma.num_vars), sess)
return np.concatenate((z_dirichlet, z_gaussian, z_invgamma), axis=1)
def log_prob_zi(self, i, z):
"""log q(z_i | lambda_i)"""
if i < self.dirichlet.num_vars:
return self.dirichlet.log_prob_zi(i, z[:, 0:self.dirichlet.num_vars])
elif i < self.dirichlet.num_vars + self.gaussian.num_vars:
i = i - self.dirichlet.num_vars
return self.gaussian.log_prob_zi(i,
z[:, self.dirichlet.num_vars:(self.dirichlet.num_vars+self.gaussian.num_vars)])
elif i < self.num_vars:
i = i - self.dirichlet.num_vars - self.gaussian.num_vars
return self.invgamma.log_prob_zi(i,
z[:, (self.dirichlet.num_vars+self.gaussian.num_vars):])
else:
raise
def __init__(self, K, D):
self.K = K
self.D = D
self.dirichlet = MFDirichlet(1, K)
self.gaussian = MFGaussian(K*D)
self.invgamma = MFInvGamma(K*D)
Likelihood.__init__(self, self.dirichlet.num_vars + \
self.gaussian.num_vars + \
self.invgamma.num_vars)
self.num_params = self.dirichlet.num_params + \
self.gaussian.num_params + \
self.invgamma.num_params