Temporal Sequence Networks

This page documents the Wave 2.1 temporal sequence backbones used for TE regression candidates.

Temporal sequence networks for Wave 2.1 TE regression candidates.

scripts.models.temporal_sequence_network.resolve_sequence_readout_tensor(sequence_tensor, readout_position)[source]

Select one timestep representation from a batched sequence tensor.

Parameters:

  • sequence_tensor (Tensor)

  • readout_position (str)

Return type:

Tensor

class scripts.models.temporal_sequence_network.TemporalConvolutionNetwork(input_size, channel_size, output_size=1, kernel_size=5, activation_name='GELU', dropout_probability=0.10, readout_position='center')[source]

Bases: Module

Causal-free temporal convolutional regressor for TE sequence windows.

Parameters:

  • input_size (int)

  • channel_size (list[int])

  • output_size (int)

  • kernel_size (int)

  • activation_name (str)

  • dropout_probability (float)

  • readout_position (str)

__init__(input_size, channel_size, output_size=1, kernel_size=5, activation_name='GELU', dropout_probability=0.10, readout_position='center')[source]

Initialize the temporal convolutional regression backbone.

Parameters:

  • input_size (int)

  • channel_size (list[int])

  • output_size (int)

  • kernel_size (int)

  • activation_name (str)

  • dropout_probability (float)

  • readout_position (str)

Return type:

None

forward(input_tensor)[source]

Run the temporal convolutional network on rank-3 input windows.

Parameters:

input_tensor (Tensor)

Return type:

Tensor

class scripts.models.temporal_sequence_network.RecurrentSequenceNetwork(recurrent_type, input_size, hidden_size, output_size=1, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center')[source]

Bases: Module

GRU or LSTM sequence regressor with an explicit temporal readout.

Parameters:

  • recurrent_type (str)

  • input_size (int)

  • hidden_size (int)

  • output_size (int)

  • num_layers (int)

  • dropout_probability (float)

  • bidirectional (bool)

  • readout_position (str)

__init__(recurrent_type, input_size, hidden_size, output_size=1, num_layers=2, dropout_probability=0.10, bidirectional=False, readout_position='center')[source]

Initialize a recurrent sequence regression backbone.

Parameters:

  • recurrent_type (str)

  • input_size (int)

  • hidden_size (int)

  • output_size (int)

  • num_layers (int)

  • dropout_probability (float)

  • bidirectional (bool)

  • readout_position (str)

Return type:

None

forward(input_tensor)[source]

Run the recurrent sequence network on rank-3 input windows.

Parameters:

input_tensor (Tensor)

Return type:

Tensor