f2222bea458b02680de06f7c2445662ca6badf16,gpytorch/models/pyro_variational_gp.py,PyroVariationalGP,model,#PyroVariationalGP#Any#Any#,27

Before Change

        num_minibatch = variational_dist_f.event_shape.numel()

        // Now make the variational distribution Normal - for conditional indepdence
        variational_dist_f = pyro.distributions.Normal(variational_dist_f.mean, variational_dist_f.variance)

        with pyro.poutine.scale(scale=float(self.num_data / num_minibatch)):
            return self.likelihood.pyro_sample_outputs(y, variational_dist_f)

After Change

        // Draw samples from p(u) for KL divergence computation
        inducing_values_samples = self.sample_inducing_values(prior_dist)
        sample_shape = inducing_values_samples.shape[:-len(prior_dist.shape())] + \
            torch.Size([1] * len(prior_dist.batch_shape))

        // Get the variational distribution for the function
        function_dist = self(input)

        // Go from function -> output
        num_minibatch = function_dist.batch_shape[-1]
        with pyro.poutine.scale(scale=float(self.num_data / num_minibatch)):
            return self.likelihood.pyro_sample_output(
                output, function_dist, *params, **kwargs, sample_shape=sample_shape
            )

    def sample_inducing_values(self, inducing_values_dist):
        
        Sample values from the inducing point distribution `p(u)` or `q(u)`.

Instances