956ac38fd58a1ef65c18dffc06a22d2e628e3a16,GPflow/conditionals.py,,gp_predict,#Any#Any#Any#Any#,103

Before Change

    
 
    //compute kernel stuff
    num_data = X.shape[0]
    Kdiag = kern.Kdiag(Xnew)
    Kmn = kern.K(X, Xnew)
    Kmm = kern.K(X) + eye(num_data)*1e-4
    Lm = tf.cholesky(Kmm)

    //this is O(N M^2)
    A = tf.user_ops.triangular_solve(Lm, Kmn, "lower")
    B = tf.user_ops.triangular_solve(tf.transpose(Lm), A, "upper") // B is Kmm^{-1} Kmn

    //construct the mean and variance of q(f*)
    fmean = tf.matmul(B.T, F)
    fvar = Kdiag[:,None] - tf.reduce_sum(tf.square(A), 0)[:,None]

    return fmean, fvar

After Change

    
 
    //compute kernel stuff
    num_data = tf.shape(X)[0]
    Kdiag = kern.Kdiag(Xnew)
    Kmn = kern.K(X, Xnew)
    Kmm = kern.K(X) + eye(num_data)*1e-4
    Lm = tf.cholesky(Kmm)

    //this is O(N M^2)
    A = tf.user_ops.triangular_solve(Lm, Kmn, "lower")
    B = tf.user_ops.triangular_solve(tf.transpose(Lm), A, "upper") // B is Kmm^{-1} Kmn

    //construct the mean and variance of q(f*)
    fmean = tf.matmul(tf.transpose(B), F)
    fvar = Kdiag - tf.reduce_sum(tf.square(A), 0)
    fvar = tf.expand_dims(fvar, 1)

    return fmean, fvar

Instances