def score(self, X, mask):
from scipy.stats import norm, multivariate_normal, gamma
if not hasattr(self, 'n_relations'):
self.n_entities, self.n_relations, _ = X.shape
score = 0.
p = self.p_weights.argmax()
for k in range(self.n_relations):
mean = np.dot(np.dot(self.E[p], self.R[p][k]), self.E[p].T)
score += np.sum(norm.logpdf(X[k].flatten(), mean.flatten(), np.sqrt(self.var_x)) * mask[k].flatten())
score += np.sum(norm.logpdf(self.R[p][k].flatten(), 0, np.sqrt(self.var_r[p])))
for i in range(self.n_entities):
score += multivariate_normal.logpdf(self.E[p][i], np.zeros(self.n_dim),
np.identity(self.n_dim) * self.var_e[p])
if self.sample_prior:
score += gamma.logpdf(self.var_e[p], loc=self.e_alpha, shape=self.e_beta)
score += gamma.logpdf(self.var_r[p], loc=self.r_alpha, shape=self.r_beta)
return score
def score(self, X):
from scipy.stats import norm, multivariate_normal
if not hasattr(self, 'n_relations'):
self.n_entities, self.n_relations, _ = X.shape
score = 0
for k in range(self.n_relations):
mean = np.dot(np.dot(self.E, self.R[k]), self.E.T)
if self.controlled_var:
score += np.sum(norm.logpdf(X[k].flatten(), mean.flatten(), np.sqrt(self.var_X[k].flatten())))
else:
score += np.sum(norm.logpdf(X[k].flatten(), mean.flatten(), np.sqrt(self.var_x)))
score += np.sum(norm.logpdf(self.R[k].flatten(), 0, np.sqrt(self.var_r)))
for i in range(self.n_entities):
score += multivariate_normal.logpdf(self.E[i], np.zeros(self.n_dim), np.identity(self.n_dim) * self.var_e)
if self.sample_prior:
score += (self.e_alpha - 1.) * np.log(self.var_e) - self.e_beta * self.var_e
score += (self.r_alpha - 1.) * np.log(self.var_r) - self.r_beta * self.var_r
return score