nonlinear.py 文件源码

def forward(self, y, weights, mean, std):
        """
        Presents a maximum a-priori objective for a set of predicted means, mixture components,
        and standard deviations to model a given ground-truth 'y'. Modeled using negative log
        likelihood.

        :param y: Non-linear target.
        :param weights: Predicted mixture components.
        :param mean: Predicted mixture means.
        :param std: Predicted mixture standard deviations.
        :return:
        """
        normalization = 1.0 / ((2.0 * math.pi) ** 0.5)
        gaussian_sample = (y.expand_as(mean) - mean) * torch.reciprocal(std)
        gaussian_sample = normalization * torch.reciprocal(std) * torch.exp(-0.5 * gaussian_sample ** 2)

        return -torch.mean(torch.log(torch.sum(weights * gaussian_sample, dim=1)))