def expected_statistics(h, b):
w_prob = unlog(h[0] + b[1])
hits_expected = fast_average(hits_pre[0], w_prob)
for k in range(1, params['M']):
w_prob_new = np.zeros(n_w[k])
for w_prev in range(n_w[k-1]):
w_weight = unlog(h[k][w_prev,:] + b[k+1])
w_prob_new += w_weight * w_prob[w_prev]
hits_expected += (w_prob[w_prev] *
fast_average(hits_pre[k][w_prev], w_weight))
w_prob = w_prob_new
return hits_expected
def grad_neg_log_likelihood(theta_vec):
theta = np.reshape(theta_vec, theta_dim)
h, b = dp(theta)
# Compute expected statistics
w_prob = unlog(h[0] + b[1])
hits_expected = fast_average(hits[0], w_prob)
for k in range(1, params['M']):
w_prob_new = np.zeros(n_w[k])
for w_prev in range(n_w[k-1]):
w_weight = unlog(h[k][w_prev,:] + b[k+1])
w_prob_new += w_weight * w_prob[w_prev]
hits_expected += (w_prob[w_prev] *
fast_average(hits[k][w_prev], w_weight))
w_prob = w_prob_new
return np.reshape(hits_expected - hits_observed, theta_vec.shape)