def _test_log_prob_zi(n_minibatch, num_factors, K):
multinomial = Multinomial(num_factors, K)
multinomial.pi = tf.constant(1.0/K, shape=[num_factors, K])
with sess.as_default():
pi = multinomial.pi.eval()
z = np.zeros((n_minibatch, K*num_factors))
for i in xrange(num_factors):
z[:, (i*K):((i+1)*K)] = np.random.multinomial(1, pi[i, :], size=n_minibatch)
z_tf = tf.constant(z, dtype=tf.float32)
for i in xrange(num_factors):
# NOTE: since Tensorflow has no special functions, the values here are
# only an approximation
assert np.allclose(
multinomial.log_prob_zi(i, z_tf).eval(),
multinomial_logpmf_vec(z[:, (i*K):((i+1)*K)], 1, pi[i, :]),
atol=1e-4)
def _test_log_prob_zi(n_minibatch, num_factors, K):
multinomial = Multinomial(num_factors, K)
multinomial.pi = tf.constant(1.0 / K, shape=[num_factors, K])
with sess.as_default():
pi = multinomial.pi.eval()
z = np.zeros((n_minibatch, K * num_factors))
for i in range(num_factors):
z[:,
(i * K):((i + 1) * K)] = np.random.multinomial(1,
pi[i, :],
size=n_minibatch)
z_tf = tf.constant(z, dtype=tf.float32)
for i in range(num_factors):
# NOTE: since Tensorflow has no special functions, the values here are
# only an approximation
assert np.allclose(multinomial.log_prob_zi(i, z_tf).eval(),
multinomial_logpmf_vec(
z[:, (i * K):((i + 1) * K)], 1, pi[i, :]),
atol=1e-4)
