Latent State Hysteresis Network

Latent-state hysteresis-aware network for Wave 4.4 TE probes.

class scripts.models.latent_state_hysteresis_network.CausalTemporalStateEncoder(input_size, hidden_size, channel_size=None, kernel_size=5, activation_name='GELU', dropout_probability=0.10)[source]

Bases: Module

Compact causal temporal convolution state encoder.

Parameters:

  • input_size (int)

  • hidden_size (int)

  • channel_size (list[int] | None)

  • kernel_size (int)

  • activation_name (str)

  • dropout_probability (float)

__init__(input_size, hidden_size, channel_size=None, kernel_size=5, activation_name='GELU', dropout_probability=0.10)[source]

Initialize the causal temporal state encoder.

Parameters:

  • input_size (int)

  • hidden_size (int)

  • channel_size (list[int] | None)

  • kernel_size (int)

  • activation_name (str)

  • dropout_probability (float)

Return type:

None

forward(input_tensor)[source]

Encode a batch-first causal sequence into one latent state.

Parameters:

input_tensor (Tensor)

Return type:

Tensor

class scripts.models.latent_state_hysteresis_network.LatentStateHysteresisNetwork(input_size, output_size=1, latent_encoder_type='gru', latent_hidden_size=96, latent_num_layers=2, latent_dropout_probability=0.10, latent_channel_size=None, latent_kernel_size=5, latent_activation_name='GELU', readout_position='last', base_hidden_size=None, head_hidden_size=None, head_activation_name='GELU', head_dropout_probability=0.05, use_layer_norm=True, offset_scale=1.0, residual_scale=1.0)[source]

Bases: Module

Causal latent-state model with base, offset, and residual TE heads.

Parameters:

  • input_size (int)

  • output_size (int)

  • latent_encoder_type (str)

  • latent_hidden_size (int)

  • latent_num_layers (int)

  • latent_dropout_probability (float)

  • latent_channel_size (list[int] | None)

  • latent_kernel_size (int)

  • latent_activation_name (str)

  • readout_position (str)

  • base_hidden_size (list[int] | None)

  • head_hidden_size (list[int] | None)

  • head_activation_name (str)

  • head_dropout_probability (float)

  • use_layer_norm (bool)

  • offset_scale (float)

  • residual_scale (float)

__init__(input_size, output_size=1, latent_encoder_type='gru', latent_hidden_size=96, latent_num_layers=2, latent_dropout_probability=0.10, latent_channel_size=None, latent_kernel_size=5, latent_activation_name='GELU', readout_position='last', base_hidden_size=None, head_hidden_size=None, head_activation_name='GELU', head_dropout_probability=0.05, use_layer_norm=True, offset_scale=1.0, residual_scale=1.0)[source]

Initialize the latent-state hysteresis-aware TE probe.

Parameters:

  • input_size (int)

  • output_size (int)

  • latent_encoder_type (str)

  • latent_hidden_size (int)

  • latent_num_layers (int)

  • latent_dropout_probability (float)

  • latent_channel_size (list[int] | None)

  • latent_kernel_size (int)

  • latent_activation_name (str)

  • readout_position (str)

  • base_hidden_size (list[int] | None)

  • head_hidden_size (list[int] | None)

  • head_activation_name (str)

  • head_dropout_probability (float)

  • use_layer_norm (bool)

  • offset_scale (float)

  • residual_scale (float)

Return type:

None

resolve_readout_feature_tensor(sequence_tensor)[source]

Extract the current operating-state feature tensor.

Parameters:

sequence_tensor (Tensor)

Return type:

Tensor

encode_latent_state(normalized_input_tensor)[source]

Encode the causal operating-history window into one latent state.

Parameters:

normalized_input_tensor (Tensor)

Return type:

Tensor

compute_auxiliary_output_dictionary(input_tensor, normalized_input_tensor)[source]

Expose base, latent, offset, residual, and final 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 context and normalized model input.

Parameters:

  • input_tensor (Tensor)

  • normalized_input_tensor (Tensor)

Return type:

Tensor

forward(normalized_input_tensor)[source]

Run inference from normalized sequence inputs.

Parameters:

normalized_input_tensor (Tensor)

Return type:

Tensor