Wave 5.1 Harmonic Prior Residual Network

Embryonic Wave 5.1 harmonic-prior residual TE model.

class scripts.models.wave3_harmonic_prior_residual_network.Wave3HarmonicPriorResidualNetwork(input_size, output_size=1, harmonic_order=240, coefficient_mode='linear_conditioned', harmonic_index_list=None, residual_hidden_size=None, residual_activation_name='GELU', residual_dropout_probability=0.05, residual_use_layer_norm=True, residual_scale=1.0, readout_position='center', freeze_structured_branch=False, low_order_harmonic_index_list=None, stable_middle_harmonic_index_list=None, high_order_harmonic_index_list=None)[source]

Bases: Module

Structured harmonic reconstruction plus residual correction skeleton.

Parameters:

  • input_size (int)

  • output_size (int)

  • harmonic_order (int)

  • coefficient_mode (str)

  • harmonic_index_list (Sequence[int] | None)

  • residual_hidden_size (list[int] | None)

  • residual_activation_name (str)

  • residual_dropout_probability (float)

  • residual_use_layer_norm (bool)

  • residual_scale (float)

  • readout_position (str)

  • freeze_structured_branch (bool)

  • low_order_harmonic_index_list (Sequence[int] | None)

  • stable_middle_harmonic_index_list (Sequence[int] | None)

  • high_order_harmonic_index_list (Sequence[int] | None)

__init__(input_size, output_size=1, harmonic_order=240, coefficient_mode='linear_conditioned', harmonic_index_list=None, residual_hidden_size=None, residual_activation_name='GELU', residual_dropout_probability=0.05, residual_use_layer_norm=True, residual_scale=1.0, readout_position='center', freeze_structured_branch=False, low_order_harmonic_index_list=None, stable_middle_harmonic_index_list=None, high_order_harmonic_index_list=None)[source]

Initialize the embryonic Wave 5.1 model.

Parameters:

  • input_size (int) – Input feature count, with angular position in the first feature column.

  • output_size (int) – Regression target count. The skeleton supports the repository’s scalar TE target.

  • harmonic_order (int) – Contiguous fallback harmonic order.

  • coefficient_mode (str) – Structured branch coefficient mode.

  • harmonic_index_list (Sequence[int] | None) – Explicit harmonic set. The current Wave 5.1 default should use the recovered paper set.

  • residual_hidden_size (list[int] | None) – Residual branch hidden widths.

  • residual_activation_name (str) – Residual branch activation name.

  • residual_dropout_probability (float) – Residual branch dropout.

  • residual_use_layer_norm (bool) – Whether the residual branch uses layer normalization.

  • residual_scale (float) – Multiplicative residual scale before summation.

  • readout_position (str) – Readout position for sequence batches.

  • freeze_structured_branch (bool) – Whether the harmonic prior is frozen.

  • low_order_harmonic_index_list (Sequence[int] | None) – Diagnostic low-order group.

  • stable_middle_harmonic_index_list (Sequence[int] | None) – Diagnostic middle group.

  • high_order_harmonic_index_list (Sequence[int] | None) – Diagnostic high-order group.

Return type:

None

resolve_readout_tensor(input_tensor)[source]

Return point-level features from point or sequence input.

Parameters:

input_tensor (Tensor)

Return type:

Tensor

compute_auxiliary_output_dictionary(input_tensor, normalized_input_tensor)[source]

Expose structured, residual, and combined prediction tensors.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

dict[str, Tensor]

forward_with_input_context(input_tensor, normalized_input_tensor)[source]

Predict normalized TE using raw angle context and normalized features.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

Tensor

forward(normalized_input_tensor)[source]

Run inference from normalized inputs when raw context is unavailable.

Parameters:

normalized_input_tensor (Tensor)

Return type:

Tensor