Cleaning up and a fix in pytorch marginal log-lik of bottleneck NNGP

import math

import torch

from torch.distributions import Normal

import torch.nn as nn

class Hyperparameters:

        self.v_w = nn.Parameter(data=torch.empty([1], dtype=self.dtype, device=self.device), requires_grad=True)

        self.v_n = nn.Parameter(data=torch.empty([1], dtype=self.dtype, device=self.device), requires_grad=True)

        self.prior = self.default_prior()

        self.samples = samples

    def default_prior(self):

        return Normal(

            torch.zeros(self.num_params(), dtype=self.dtype, device=self.device),

            torch.ones(self.num_params(), dtype=self.dtype, device=self.device)

        )

    def num_params(self):

        """ Get the number of model parameters. """

        return sum(p.numel() for p in self.parameters())

    def log_lik(self, x, y, samples=None):

        """ Log-likelihood """

        gram = (x @ x.t()).unsqueeze(0)

        grams = (x @ x.t()).unsqueeze(0)

        for (depth, width) in zip(self.hp.depths[:-1], self.hp.widths):

            Ks = self.K(gram, depth, x.shape[0], self.hp.noise)

            Ks = self.K(grams, depth, x.shape[0], self.hp.noise)

            Ls = torch.stack([torch.cholesky(K) for K in Ks], 0)

            if samples is None: