pyshred.models.sequence_models package#

Submodules#

pyshred.models.sequence_models.abstract_sequence module#

class pyshred.models.sequence_models.abstract_sequence.AbstractSequence[source]#

Bases: ABC, Module

Abstract base class for all sequence models.

abstractmethod forward(x)[source]#

Forward pass through the sequence model.

abstractmethod initialize(input_size)[source]#
abstract property model_name#

Returns the name of the sequence model.

Returns:#

str

The name of the model.

pyshred.models.sequence_models.gru_model module#

class pyshred.models.sequence_models.gru_model.GRU(hidden_size: int = 3, num_layers: int = 1, layer_norm: bool = False)[source]#

Bases: AbstractSequence

GRU (Gated Recurrent Unit) sequence model for encoding temporal sensor dynamics.

A recurrent neural network that processes sensor measurement sequences to learn latent representations of the underlying dynamics. Supports optional layer normalization for improved training stability.

Parameters:

  • hidden_size (int, optional) – Size of the GRU hidden state. Defaults to 3.

  • num_layers (int, optional) – Number of GRU layers. Defaults to 1.

  • layer_norm (bool, optional) – Whether to apply layer normalization. Defaults to False.

Variables:

  • hidden_size (int) – Size of the GRU hidden state.

  • num_layers (int) – Number of GRU layers.

  • use_layer_norm (bool) – Whether layer normalization is applied.

  • output_size (int) – Size of the output (equals hidden_size).

forward(x)[source]#

Forward pass through the GRU model.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch_size, sequence_length, input_size).

Returns:

Output tensor with latent representations. Shape depends on decoder type: - MLP decoder: (batch_size, hidden_size) - UNET decoder: (batch_size, hidden_size, sequence_length)

Return type:

torch.Tensor

Raises:

TypeError – If decoder type is not supported.

initialize(input_size: int, decoder_type, **kwargs)[source]#

Initialize the GRU with input size and decoder.

Parameters:

  • input_size (int) – Number of input features (sensor measurements).

  • decoder_type (str) – Decoder model type.

  • **kwargs – Additional keyword arguments.

property model_name#

Name of the sequence model.

Returns:

Returns “GRU”.

Return type:

str

pyshred.models.sequence_models.lstm_model module#

class pyshred.models.sequence_models.lstm_model.LSTM(hidden_size: int = 64, num_layers: int = 2, layer_norm: bool = False)[source]#

Bases: AbstractSequence

LSTM sequence model for encoding temporal sensor dynamics.

Parameters:

  • hidden_size (int, optional) – Size of the hidden state. Defaults to 64.

  • num_layers (int, optional) – Number of LSTM layers. Defaults to 2.

  • layer_norm (bool, optional) – Whether to apply layer normalization. Defaults to False.

Variables:

  • hidden_size (int) – Size of the hidden state.

  • num_layers (int) – Number of LSTM layers.

  • use_layer_norm (bool) – Whether layer normalization is applied.

forward(x)[source]#

Forward pass through the LSTM model.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch_size, sequence_length, input_size).

Returns:

Output tensor with latent representations.

Return type:

torch.Tensor

initialize(input_size: int, decoder_type, **kwargs)[source]#

Initialize the LSTM with input size and decoder.

Parameters:

  • input_size (int) – Number of input features.

  • decoder_type (str) – Decoder model type.

  • **kwargs – Additional keyword arguments.

property model_name#

Name of the sequence model.

Returns:

Returns “LSTM”.

Return type:

str

pyshred.models.sequence_models.transformer_model module#

class pyshred.models.sequence_models.transformer_model.PositionalEncoding(d_model, max_sequence_length=5000, dropout=0.1)[source]#

Bases: Module

Sinusoidal positional encoding for Transformer models.

Adds learnable positional information to input embeddings to help the model understand sequence order.

Parameters:

  • d_model (int) – Dimensionality of the model embeddings.

  • max_sequence_length (int, optional) – Maximum sequence length to precompute encodings for. Defaults to 5000.

  • dropout (float, optional) – Dropout probability. Defaults to 0.1.

Variables:

pe (torch.Tensor) – Precomputed positional encoding tensor.

forward(x)[source]#

Add positional encoding to input embeddings.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch_size, seq_len, d_model).

Returns:

Input with positional encoding added, same shape as input.

Return type:

torch.Tensor

class pyshred.models.sequence_models.transformer_model.TRANSFORMER(d_model: int = 128, nhead: int = 16, dropout: float = 0.2, layer_norm: bool = False)[source]#

Bases: AbstractSequence

Transformer-based sequence model for encoding temporal sensor dynamics.

Uses self-attention mechanisms to capture long-range dependencies in sensor measurement sequences. Includes positional encoding and supports optional layer normalization.

Parameters:

  • d_model (int, optional) – Dimensionality of the model (embedding size). Defaults to 128.

  • nhead (int, optional) – Number of attention heads. Defaults to 16.

  • dropout (float, optional) – Dropout probability. Defaults to 0.2.

  • layer_norm (bool, optional) – Whether to apply layer normalization. Defaults to False.

Variables:

  • d_model (int) – Model dimensionality.

  • hidden_size (int) – Hidden size (same as d_model).

  • output_size (int) – Output size (same as d_model).

  • use_layer_norm (bool) – Whether layer normalization is applied.

forward(x: Tensor) → Tensor[source]#

Forward pass through the Transformer model.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch_size, sequence_length, input_size).

Returns:

Output tensor with latent representations. Shape depends on decoder type: - MLP decoder: (batch_size, d_model) - UNET decoder: (batch_size, d_model, sequence_length)

Return type:

torch.Tensor

Raises:

TypeError – If decoder type is not supported.

initialize(input_size: int, lags: int, decoder_type, **kwargs)[source]#

Initialize the Transformer with input size, sequence length, and decoder.

Parameters:

  • input_size (int) – Number of input features (sensor measurements).

  • lags (int) – Length of input sequences.

  • decoder_type (str) – Decoder model type.

  • **kwargs – Additional keyword arguments.

property model_name#

Name of the sequence model.

Returns:

Returns “Transformer”.

Return type:

str

Module contents#