bbc4e3ee65ab504be761e3b5487ce4ccdd589da0,GPflow/svgp.py,SVGP,build_likelihood,#SVGP#,46

Before Change

                def f(w):
                    R = tf.triu(w)
                    return -tf.reduce_sum(tf.log(tf.user_ops.get_diag(R))) + 0.5*tf.reduce_sum(tf.square(R))
                KL += theano.scan(f, self.q_sqrt.swapaxes(0,2))[0].sum()
            fmean, fvar = conditionals.gaussian_gp_predict_whitened(self.X, self.Z, self.kern, self.q_mu, self.q_sqrt)
        else:
            L = tf.cholesky(self.kern.K(self.Z) + eye(self.num_inducing) * 1e-4)

After Change

                KL += -tf.reduce_sum(tf.log(self.q_sqrt)) + 0.5*tf.reduce_sum(tf.square(self.q_sqrt))
            else:
                //here we loop through all the independent columns, extracting the triangular part. 
                for d in range(self.num_latent):
                    L = tf.user_ops.triangle(self.q_sqrt[d,:,:], "lower")
                    KL += -tf.reduce_sum(tf.log(tf.user_ops.get_diag(L))) + 0.5*tf.reduce_sum(tf.square(L))
            fmean, fvar = conditionals.gaussian_gp_predict_whitened(self.X, self.Z, self.kern, self.q_mu, self.q_sqrt)
        else:
            L = tf.cholesky(self.kern.K(self.Z) + eye(self.num_inducing) * 1e-4)
            alpha = tf.user_ops.triangular_solve(L, self.q_mu, "lower")

Instances