Periodic Temporal Sequence Networks

This page documents the Wave 2.2 harmonic-temporal hybrid sequence backbones used for TE regression candidates.

Periodic temporal sequence networks for harmonic-aware TE windows.

class scripts.models.periodic_temporal_sequence_network.PeriodicTemporalSequenceNetwork(temporal_model_type, input_size, output_size=1, harmonic_order=8, harmonic_index_list=None, include_raw_angle_feature=True, channel_size=None, kernel_size=5, activation_name='GELU', hidden_size=128, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center')[source]

Bases: Module

Temporal TE sequence model with per-timestep harmonic angle features.

Parameters:

  • temporal_model_type (str)

  • input_size (int)

  • output_size (int)

  • harmonic_order (int)

  • harmonic_index_list (list[int] | None)

  • include_raw_angle_feature (bool)

  • channel_size (list[int] | None)

  • kernel_size (int)

  • activation_name (str)

  • hidden_size (int)

  • num_layers (int)

  • dropout_probability (float)

  • bidirectional (bool)

  • readout_position (str)

__init__(temporal_model_type, input_size, output_size=1, harmonic_order=8, harmonic_index_list=None, include_raw_angle_feature=True, channel_size=None, kernel_size=5, activation_name='GELU', hidden_size=128, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center')[source]

Initialize one periodic temporal TE sequence backbone.

Parameters:

  • temporal_model_type (str) – Temporal backbone selector. Supported values are temporal_convolution, 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.

  • harmonic_index_list (list[int] | None) – Optional explicit non-negative harmonic list. Positive indices create sine/cosine pairs and 0 follows the existing DC/bias convention.

  • include_raw_angle_feature (bool) – Whether to keep the normalized raw angle alongside the harmonic feature expansion.

  • channel_size (list[int] | None) – Temporal convolution channel widths.

  • kernel_size (int) – Temporal convolution kernel size.

  • activation_name (str) – Temporal convolution activation name.

  • hidden_size (int) – Recurrent hidden size for GRU and LSTM.

  • num_layers (int) – Recurrent layer count.

  • dropout_probability (float) – Dropout probability used by the temporal backbone.

  • bidirectional (bool) – Whether recurrent backbones are bidirectional.

  • readout_position (str) – Sequence readout position passed to the temporal backbone.

Return type:

None

build_periodic_feature_tensor(angular_position_deg)[source]

Build sine/cosine harmonic features for rank-2 or rank-3 angles.

Parameters:

angular_position_deg (Tensor)

Return type:

Tensor

build_expanded_sequence_tensor(input_tensor, normalized_input_tensor)[source]

Build the expanded per-timestep feature tensor for the backbone.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

Tensor

forward_with_input_context(input_tensor, normalized_input_tensor)[source]

Predict normalized TE from harmonic-aware sequence windows.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

Tensor