shredx.modules.sindy_layer.SINDyLayer

class shredx.modules.sindy_layer.SINDyLayer(hidden_size: int, forecast_length: int, device: str = 'cpu', strict_symmetry: bool = True, std_init: float = 0.1, **kwargs)

Bases: Module

Differentiable SINDy layer for ODE-based forecasting.

Learns sparse polynomial dynamics from data and uses ODE integration for arbitrary-length forecasting. Supports both strict symmetry (parameterized via lower triangle) and general coefficient matrices.

Parameters:

hidden_sizeint

Input/output dimension of the layer.

forecast_lengthint

Number of future timesteps to predict.

devicestr, optional

Device to place the model on. Default is "cpu".

strict_symmetrybool, optional

If True, enforces symmetric coefficient matrix via lower triangle parameterization. Default is True.

std_initfloat, optional

Standard deviation for initial coefficients. Default is 0.1.

**kwargs

Additional keyword arguments (ignored).

Methods

dense_matrix_from_symmetric_params(params)	Build symmetric dense matrix from lower-triangle parameters.
forward(x)	Integrate learned ODE dynamics to produce multi-step forecasts.
get_dense_sindy_coefficients()	Return dense SINDy coefficient matrix.
get_eigenvalues()	Return eigenvalues of the SINDy coefficient matrix (1j * matrix).
get_raw_sindy_coefficients()	Return raw SINDy coefficients (1D or full matrix).
set_raw_sindy_coefficients(coefficients)	Set raw SINDy coefficients from a tensor.

Notes

Class Methods:

get_dense_sindy_coefficients():

• Converts symmetric parameters (1D) to a dense matrix when strict_symmetry is True; otherwise returns the stored full matrix.

• Returns:

  • Float[torch.Tensor, "hidden_size hidden_size"]. Dense SINDy coefficient matrix.

get_raw_sindy_coefficients():

• Returns raw SINDy coefficients (1D lower triangle or full matrix).

• Returns:

  • Float[torch.Tensor, "library_dim library_dim"] when strict_symmetry is False, else Float[torch.Tensor, "num_params"] (lower triangle 1D).

set_raw_sindy_coefficients(coefficients):

• Updates the layer parameters in-place from raw SINDy coefficients.

• Parameters:

  • coefficients : Float[torch.Tensor, "library_dim library_dim"] | Float[torch.Tensor, "num_params"]. Raw SINDy coefficients (same shape as get_raw_sindy_coefficients).

• Returns:

  • None.

dense_matrix_from_symmetric_params(params):

• Builds a symmetric matrix by filling the lower triangle and reflecting.

• Parameters:

  • params : Float[torch.Tensor, "num_params"]. 1D lower-triangle coefficients.

• Returns:

  • Float[torch.Tensor, "library_dim library_dim"]. Dense symmetric matrix.

get_eigenvalues():

• Returns eigenvalues of the SINDy coefficient matrix (1j * matrix).

• Returns:

  • Float[torch.Tensor, "hidden_size"]. Eigenvalues.

forward(x):

• Integrates the learned ODE (dopri5) over forecast_length steps to produce multi-step forecasts.

• Parameters:

  • x : Float[torch.Tensor, "batch_size hidden_size"]. Input state.

• Returns:

  • Float[torch.Tensor, "batch_size forecast_length hidden_size"]. Rollout of predicted states.