osl_dynamics.models.obs_mod.set_covariances_regularizer#
- osl_dynamics.models.obs_mod.set_covariances_regularizer(model, range_, epsilon, scale_factor, diagonal=False, layer_name='covs')#
Set the covariances regularizer based on training data.
If config.diagonal_covariances is True, a log-normal prior is applied to the diagonal of the covariance matrices with
mu=0,sigma=sqrt(log(2*range)). Otherwise, an inverse Wishart prior is applied to the covariance matrices withnu=n_channels-1+0.1,psi=diag(1/range).x- Parameters:
- modelosl_dynamics.models.*.Model.model
The model.
- range_np.ndarray
Range (max-min) of the training data for each channel. Shape is (n_channels,).
- epsilonfloat
Error added to the covariance matrices.
- scale_factorfloat
Scale factor for regularization.
- diagonalbool, optional
Whether the covariances are diagonal.
- layer_namestr, optional
Layer name of the covariances. Can be
"covs"or"group_covs".