Residual Harmonic Temporal Sequence Networks

This page documents the Wave 2.3 residual harmonic temporal sequence backbones used for TE regression candidates.

Residual harmonic temporal sequence networks for Wave 2.3 TE candidates.

class scripts.models.residual_harmonic_temporal_sequence_network.ResidualHarmonicTemporalSequenceNetwork(temporal_model_type, input_size, output_size=1, harmonic_order=12, coefficient_mode='static', harmonic_index_list=None, hidden_size=128, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center', freeze_structured_branch=False)[source]

Bases: Module

Sequence TE model with harmonic base prediction plus temporal residual.

Parameters:

  • temporal_model_type (str)

  • input_size (int)

  • output_size (int)

  • harmonic_order (int)

  • coefficient_mode (str)

  • harmonic_index_list (list[int] | None)

  • hidden_size (int)

  • num_layers (int)

  • dropout_probability (float)

  • bidirectional (bool)

  • readout_position (str)

  • freeze_structured_branch (bool)

__init__(temporal_model_type, input_size, output_size=1, harmonic_order=12, coefficient_mode='static', harmonic_index_list=None, hidden_size=128, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center', freeze_structured_branch=False)[source]

Initialize one residual harmonic recurrent sequence model.

Parameters:

  • temporal_model_type (str) – Temporal residual selector. Supported values are gru_sequence and lstm_sequence.

  • input_size (int) – Raw sequence feature count, including angular position as the first feature.

  • output_size (int) – Regression target count.

  • harmonic_order (int) – Contiguous harmonic order used when no explicit harmonic index list is provided.

  • coefficient_mode (str) – Harmonic coefficient parameterization mode passed to the structured branch.

  • harmonic_index_list (list[int] | None) – Optional explicit non-negative harmonic list.

  • hidden_size (int) – Recurrent hidden size for the residual branch.

  • num_layers (int) – Recurrent layer count.

  • dropout_probability (float) – Dropout probability used by the recurrent residual branch.

  • bidirectional (bool) – Whether the recurrent residual branch is bidirectional.

  • readout_position (str) – Sequence readout position used by both the residual branch and center-vector extraction.

  • freeze_structured_branch (bool) – Whether to freeze the structured harmonic branch parameters during optimization.

Return type:

None

resolve_readout_feature_tensor(sequence_tensor)[source]

Extract the rank-2 feature tensor used by the structured branch.

Parameters:

sequence_tensor (Tensor)

Return type:

Tensor

compute_auxiliary_output_dictionary(input_tensor, normalized_input_tensor)[source]

Expose branch-level outputs for diagnostics and metric logging.

Parameters:

  • input_tensor (Tensor) – Raw rank-3 sequence tensor whose first feature is physical angular position in degrees.

  • normalized_input_tensor (Tensor) – Normalized rank-3 sequence tensor used by the temporal residual branch.

Returns:

Structured branch output, residual branch output, and final combined prediction tensor.

Return type:

dict[str, torch.Tensor]

forward_with_input_context(input_tensor, normalized_input_tensor)[source]

Predict TE from a harmonic base plus recurrent residual correction.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

Tensor