shredx.modules.moe_mixin.MOESINDyLayerHelpersMixin

class shredx.modules.moe_mixin.MOESINDyLayerHelpersMixin

Bases: object

Mixin providing helper methods for Mixture of Experts models with SINDy layers.

Provides common functionality for printing, modifying, and analyzing SINDy coefficients across multiple expert networks.

Methods

get_sindy_layer_coefficients_eigenvalues()	Get the eigenvalues of SINDy coefficient matrices for all SINDy experts.
get_sindy_layer_coefficients_sum()	Compute the sum of absolute SINDy coefficients across all experts (sparsity regularization).
print_sindy_layer_coefficients()	Print the SINDy coefficients for all experts in a human-readable format.
set_forecast_length(forecast_length)	Set the forecast length for the model and all expert SINDy layers.
threshold_sindy_layer_coefficients(threshold)	Apply sparsity thresholding to SINDy coefficients for all experts.

Notes

Class Methods:

print_sindy_layer_coefficients():

• Prints the SINDy coefficients for all experts in a human-readable format (polynomial expression per hidden dimension).

• Returns:

  • None.

set_forecast_length(forecast_length):

• Sets the forecast length for the model and all expert SINDy layers.

• Parameters:

  • forecast_length : int. Number of timesteps to forecast.

• Returns:

  • None.

get_sindy_layer_coefficients_eigenvalues():

• Returns the eigenvalues of SINDy coefficient matrices for all experts.

• Returns:

  • list. List of eigenvalue tensors, one per expert.

get_sindy_layer_coefficients_sum():

• Computes the sum of absolute SINDy coefficients across all experts (used as a sparsity regularization term).

• Returns:

  • float. Sum of square roots of absolute coefficient sums.

threshold_sindy_layer_coefficients(threshold, verbose):

• Applies sparsity thresholding: sets coefficients with absolute value below the threshold to zero for all experts.

• Parameters:

  • threshold : float. Coefficients below this are zeroed.

  • verbose : bool, optional. If True, print information about thresholding. Default is False.

• Returns:

  • None.