a9273fb766562261b2ccedb635c71363fb10f06c,GPflow/svgp.py,SVGP,build_likelihood,#SVGP#,46

Before Change

                    R = tf.triu(w)
                    S_q = tf.matmul(tf.transpose(R), R)
                    return -tf.reduce_sum(tf.log(tf.user_ops.get_diag(R))) + 0.5*tf.reduce_sum(S_q * K_inv)
                KL += theano.scan(f, self.q_sqrt.swapaxes(0,2))[].sum()
            fmean, fvar = conditionals.gaussian_gp_predict(self.X, self.Z, self.kern, self.q_mu, self.q_sqrt)

After Change

                KL += 0.5 * tf.reduce_sum(tf.expand_dims(tf.user_ops.get_diag(K_inv), 1) * tf.square(self.q_sqrt))
            else:
                for d in range(self.num_latent):
                    L = tf.user_ops.triangle(self.q_sqrt[d,:,:], "lower")
                    S = tf.matmul(L, tf.transpose(L))
                    KL += -tf.reduce_sum(tf.log(tf.user_ops.get_diag(L))) + 0.5*tf.reduce_sum(S * K_inv)
            fmean, fvar = conditionals.gaussian_gp_predict(self.X, self.Z, self.kern, self.q_mu, self.q_sqrt)

Instances