Method grad

numpy_ml/neural_nets/losses/losses.py:296–328

Compute the gradient of the VLB with regard to the network parameters. Parameters ---------- y : :py:class:`ndarray ` of shape `(n_ex, N)` The original images. y_pred : :py:class:`ndarray ` of shape `(n_ex, N)`

(y, y_pred, t_mean, t_log_var)

    @staticmethod
    def grad(y, y_pred, t_mean, t_log_var):
        """
        Compute the gradient of the VLB with regard to the network parameters.

        Parameters
        ----------
        y : :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, N)`
            The original images.
        y_pred : :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, N)`
            The VAE reconstruction of the images.
        t_mean: :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, T)`
            Mean of the variational distribution :math:`q(t | x)`.
        t_log_var: :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, T)`
            Log of the variance vector of the variational distribution
            :math:`q(t | x)`.

        Returns
        -------
        dY_pred : :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, N)`
            The gradient of the VLB with regard to `y_pred`.
        dLogVar : :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, T)`
            The gradient of the VLB with regard to `t_log_var`.
        dMean : :py:class:`ndarray <numpy.ndarray>` of shape `(n_ex, T)`
            The gradient of the VLB with regard to `t_mean`.
        """
        N = y.shape[0]
        eps = np.finfo(float).eps
        y_pred = np.clip(y_pred, eps, 1 - eps)

        dY_pred = -y / (N * y_pred) - (y - 1) / (N - N * y_pred)
        dLogVar = (np.exp(t_log_var) - 1) / (2 * N)
        dMean = t_mean / N
        return dY_pred, dLogVar, dMean


class WGAN_GPLoss(ObjectiveBase):