5184f1fc798a7728dd574ae9a351e16869d9ee7b,pyro/contrib/timeseries/lgssmgp.py,GenericLGSSMWithGPNoiseModel,_forecast,#GenericLGSSMWithGPNoiseModel#Any#Any#Any#,134

Before Change

        // next compute the contribution from process noise that is injected at each timestep.
        // (we need to do a cumulative sum to integrate across time for the z-state contribution)
        eye = torch.eye(self.state_dim, device=fs_cov.device, dtype=fs_cov.dtype)
        z_process_covar = self.log_trans_noise_scale_sq.exp() * eye
        N_trans_obs_shift = torch.cat([self.z_obs_matrix.unsqueeze(0), N_trans_obs[0:-1]])
        predicted_covar2z = torch.matmul(N_trans_obs_shift.transpose(-1, -2),
                                         torch.matmul(z_process_covar, N_trans_obs_shift))  // N O O

        predicted_covar = predicted_covar1z + predicted_covar1gp + gp_process_covar + \
            torch.cumsum(predicted_covar2z, dim=0)

        if include_observation_noise:
            eye = torch.eye(self.obs_dim, device=fs_cov.device, dtype=fs_cov.dtype)
            predicted_covar = predicted_covar + self._get_obs_noise_scale().pow(2.0) * eye

        return predicted_mean, predicted_covar

    def forecast(self, targets, N_timesteps):

After Change

            torch.cumsum(predicted_covar2z, dim=0)

        if include_observation_noise:
            predicted_covar = predicted_covar + self.obs_noise_scale.pow(2.0).diag_embed()

        return predicted_mean, predicted_covar

    @pyro_method

Instances