Diffusion prior

This section contains some utilties for defining the of the observations

Diffusion prior base class

class svise.sde_learning.DiffusionPrior(d: int, Q: Tensor)[source]

Abstract base class for diffusion prior. This class assumes that diffusion is constant.

Parameters:

d (int) – number of states
Q (Tensor) – Diffusion matrix init

abstract property kl_divergence: Tensor

Returns the kl-divergence between the approx. posterior and the prior

Returns:: kl-div between approx. posterior and the prior
Return type:: Tensor

abstract property process_noise: Tensor

Returns the full process nosie matrix (L Q L^T )

Returns:: full process noise matrix
Return type:: Tensor

abstract resample_weights() → None[source]: Uses reparam. trick to update any parameters if applicable.

abstract solve_linear(r: Tensor) → Tensor[source]

Solves the linear system L Q L^T x = r: r (Tensor): residual

Returns:: x = (LQL^T)^{-1} r
Return type:: Tensor

Diagonal diffusion prior base class

class svise.sde_learning.DiagonalDiffusionPrior(d: int, Q: Tensor)[source]

Base class for the case that the diffusion matrix is diagonal.

Parameters:

d (int) – number of states
Q (Tensor) – Diffusion matrix init

abstract property process_noise_diag: Tensor

Returns the diagonal elements of the process noise term

Returns:: returns diagonal element of the process noise term
Return type:: Tensor

Constant diagonal diffusion prior

class svise.sde_learning.ConstantDiagonalDiffusionPrior(d: int, Q_diag: Tensor)[source]

Class for the case the diffusion matrix is known and constant (i.e. won’t be tuned during training):

Parameters:

d (int) – number of states
Q_diag (Tensor) – diagonal component of diffusion matrix

property kl_divergence: Tensor

Returns the kl-divergence between the approx. posterior and the prior

Returns:: kl-div between approx. posterior and the prior
Return type:: Tensor

property process_noise: Tensor

Returns the full process nosie matrix (L Q L^T )

Returns:: full process noise matrix
Return type:: Tensor

property process_noise_diag: Tensor

Returns the diagonal elements of the process noise term

Returns:: returns diagonal element of the process noise term
Return type:: Tensor

resample_weights() → None[source]: Uses reparam. trick to update any parameters if applicable.

solve_linear(r: Tensor) → Tensor[source]

Solves the linear system L Q L^T x = r: r (Tensor): residual

Returns:: x = (LQL^T)^{-1} r
Return type:: Tensor

Sparse diagonal diffusion prior

class svise.sde_learning.SparseDiagonalDiffusionPrior(d: int, Q_diag: Tensor, n_reparam_samples: int, tau: float)[source]

Prior over the diffusion matrix is a sparse diagonal matrix.

Parameters:

d (int) – number of states
Q_diag (Tensor) – the starting value of the dispersion matrix
n_reparam_samples (int) – number of reparametrization samples to use
tau (float) – global scaling parameter

property kl_divergence: Tensor

Returns the kl-divergence between the approx. posterior and the prior

Returns:: kl-div between approx. posterior and the prior
Return type:: Tensor

property process_noise: Tensor

Returns the full process nosie matrix (L Q L^T )

Returns:: full process noise matrix
Return type:: Tensor

property process_noise_diag: Tensor

Returns the diagonal elements of the process noise term

Returns:: returns diagonal element of the process noise term
Return type:: Tensor

resample_weights() → None[source]: Uses reparam. trick to update any parameters if applicable.

solve_linear(r: Tensor) → Tensor[source]

Solves the linear system L Q L^T x = r: r (Tensor): residual

Returns:: x = (LQL^T)^{-1} r
Return type:: Tensor