metrics.py 文件源码

def _compute_gaussian_kernel(histogram_shape, relative_bw):
    """Compute a gaussian kernel double the size of the histogram matrix"""
    if len(histogram_shape) == 2:
        kernel_shape = [2 * n for n in histogram_shape]
        # Create a scaled grid in which the kernel is symmetric to avoid matrix
        # inversion problems when the bandwiths are very different
        bw_ratio = relative_bw[0] / relative_bw[1]
        bw = relative_bw[0]
        X, Y = np.mgrid[-bw_ratio:bw_ratio:kernel_shape[0] * 1j,
                        -1:1:kernel_shape[1] * 1j]
        grid_points = np.vstack([X.ravel(), Y.ravel()]).T
        Cov = np.array(((bw, 0), (0, bw)))**2
        K = stats.multivariate_normal.pdf(grid_points, mean=(0, 0), cov=Cov)

        return K.reshape(kernel_shape)
    else:
        grid = np.mgrid[-1:1:histogram_shape[0] * 2j]
        return stats.norm.pdf(grid, loc=0, scale=relative_bw)