16914cec3cababd56e18ac775339469f35bf4fa6,GPflow/conditionals.py,,conditional,#,5

Before Change

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

    // Compute the projection matrix A
    A = tf.matrix_triangular_solve(Lm, Kmn, lower=True)

    // compute the covariance due to the conditioning
    if full_cov:
        fvar = kern.K(Xnew) - tf.matmul(tf.transpose(A), A)
        fvar = tf.tile(tf.expand_dims(fvar, 2), [1, 1, num_columns])
    else:
        fvar = kern.Kdiag(Xnew) - tf.reduce_sum(tf.square(A), 0)
        fvar = tf.tile(tf.expand_dims(fvar, 1), [1, num_columns])

    // another backsubstitution in the unwhitened case
    if not whiten:
        A = tf.matrix_triangular_solve(tf.transpose(Lm), A, lower=False)

    // construct the conditional mean
    fmean = tf.matmul(tf.transpose(A), f)

    // add extra projected variance from q(f) if needed
    if q_sqrt is not None:
        projected_var = []
        for d in range(num_columns):
            if q_sqrt.get_shape().ndims == 2:
                LTA = A * q_sqrt[:, d:d + 1]
            elif q_sqrt.get_shape().ndims == 3:
                L = tf.user_ops.triangle(q_sqrt[:, :, d], "lower")
                LTA = tf.matmul(tf.transpose(L), A)
            else:  // pragma no cover
                raise ValueError("Bad dimension for q_sqrt: %s" %
                                 str(q_sqrt.get_shape().ndims))

After Change

    // compute kernel stuff
    num_data = tf.shape(X)[0]
    Kmn = kern.K(X, Xnew)
    Kmm = kern.K(X) + eye(num_data) * 1e-6
    Lm = tf.cholesky(Kmm)

    // Compute the projection matrix A
    A = tf.matrix_triangular_solve(Lm, Kmn, lower=True)

    // compute the covariance due to the conditioning
    if full_cov:
        fvar = kern.K(Xnew) - tf.matmul(tf.transpose(A), A)
        fvar = tf.tile(tf.expand_dims(fvar, 2), [1, 1, num_columns])
    else:
        fvar = kern.Kdiag(Xnew) - tf.reduce_sum(tf.square(A), 0)
        fvar = tf.tile(tf.expand_dims(fvar, 1), [1, num_columns])

    // another backsubstitution in the unwhitened case
    if not whiten:
        A = tf.matrix_triangular_solve(tf.transpose(Lm), A, lower=False)

    // construct the conditional mean
    fmean = tf.matmul(tf.transpose(A), f)

    // add extra projected variance from q(f) if needed
    if q_sqrt is not None:
        projected_var = []
        for d in range(num_columns):
            if q_sqrt.get_shape().ndims == 2:
                LTA = A * q_sqrt[:, d:d + 1]
            elif q_sqrt.get_shape().ndims == 3:
                L = tf.user_ops.triangle(q_sqrt[:, :, d], "lower")
                LTA = tf.matmul(tf.transpose(L), A)
            else:  // pragma no cover
                raise ValueError("Bad dimension for q_sqrt: %s" %
                                 str(q_sqrt.get_shape().ndims))

Instances