def sandwich_kernel()

in orbit/utils/kernels.py [0:0]

• 17 lines of code
• 2 McCabe index (conditional complexity)

def sandwich_kernel(x, x_i):
    """
    Parameters
    ----------
    x : array-like
        points required to compute kernel weight
    x_i : array-like
        reference points location used to compute correspondent distance of each entry points

    Returns
    -------
    np.ndarray :
        2D array with size N x M such that
        N as the number of entry points
        M as the number of reference points
        matrix entries hold the value of weight of each element

    See Also
    --------
    1. https://mc-stan.org/docs/2_24/stan-users-guide/gaussian-process-regression.html
    2. https://en.wikipedia.org/wiki/Local_regression
    """
    N = len(x)
    M = len(x_i)
    k = np.zeros((N, M), dtype=np.double)

    np_idx = np.where(x < x_i[0])
    k[np_idx, 0] = 1

    for m in range(M - 1):
        np_idx = np.where(np.logical_and(x >= x_i[m], x < x_i[m + 1]))
        total_dist = x_i[m + 1] - x_i[m]
        backward_dist = x[np_idx] - x_i[m]
        forward_dist = x_i[m + 1] - x[np_idx]
        k[np_idx, m] = forward_dist / total_dist
        k[np_idx, m + 1] = backward_dist / total_dist

    np_idx = np.where(x >= x_i[M - 1])
    k[np_idx, M - 1] = 1

    # TODO: it is probably not needed
    k = k / np.sum(k, axis=1, keepdims=True)

    return k