Recurrent Models#

RNN-based architectures for sequential time-series modeling.

RecurrentAutoEncoder#

Multi-layer recurrent autoencoder supporting LSTM, GRU, and vanilla RNN cells. Uses time-reversed encoder output for sequence reconstruction.

Recurrent autoencoder based on PyLink88/Recurrent-Autoencoder.

class chronocratic.models.recurrent.recurrentae.model.RecurrentAutoEncoder(input_dims: int, layers: tuple[int, ...], recurrent_cell_type: RecurrentCellType = RecurrentCellType.LSTM, dropout: float | tuple[float, ...] = 0.2, loss: ReconstructionLoss = ReconstructionLoss.MSE, optimizer: OptimizerName = OptimizerName.ADAM, learning_rate: float = 0.001, sync_dist: bool = False)#

Bases: LightningModule, BasicEncodingMixin

Recurrent autoencoder for time series representation learning.

Architecture based on PyLink88/Recurrent-Autoencoder. Supports LSTM, GRU, and RNN variants selected via recurrent_cell_type. The encoder maps the input to a latent sequence via stacked RNN layers; the decoder reconstructs the original sequence from the time-reversed latent sequence using a mirrored RNN stack followed by a linear projection.

Parameters:

  • input_dims – Number of input features (channels) per timestep.

  • layers – Hidden sizes for each encoder RNN layer, e.g. (64, 32). The decoder uses the reversed order.

  • recurrent_cell_type – RNN variant — LSTM, GRU, or RNN.

  • dropout – Dropout probability applied between successive layers. A single float applies uniformly; a tuple must match len(layers).

  • loss – Reconstruction objective — 'mse' or 'mae'.

  • optimizer – Optimizer — 'adam', 'adamw', or 'radam'.

  • learning_rate – Base learning rate for the optimizer.

  • sync_dist – Whether to sync logged metrics across devices.

property encoder: Module#

The encoder nn.Module.

property decoder: Module#

The decoder nn.Module.

forward(x: Tensor) → Tensor#

Encode x, reverse the latent sequence, and reconstruct.

training_step(batch: Tensor, _batch_idx: int) → Tensor#

Compute and log reconstruction loss for a training batch.

validation_step(batch: Tensor, _batch_idx: int) → Tensor#

Compute and log reconstruction loss for a validation batch.

configure_optimizers() → OptimizerLRScheduler#

Return the optimizer configured with the model’s learning rate.

Configuration for the RecurrentAutoEncoder model.

class chronocratic.models.recurrent.recurrentae.config.RecurrentAutoEncoderModelParameters(*, input_dims: int, layers: tuple[int, ...], recurrent_cell_type: RecurrentCellType = RecurrentCellType.LSTM, dropout: float | tuple[float, ...] = 0.2, loss: ReconstructionLoss = ReconstructionLoss.MSE, optimizer: OptimizerName = OptimizerName.ADAM, learning_rate: float = 0.001, sync_dist: bool = False)#

Bases: object

Configuration for the recurrent autoencoder model.

Parameters:

  • input_dims – Number of input features (channels) per timestep.

  • layers – Hidden sizes for each encoder RNN layer, e.g. (64, 32). The decoder uses the reversed order.

  • recurrent_cell_type – RNN variant — LSTM, GRU, or RNN.

  • dropout – Dropout probability applied between successive layers. A single float applies uniformly; a tuple must match len(layers).

  • loss – Reconstruction objective — 'mse' or 'mae'.

  • optimizer – Optimizer — 'adam', 'adamw', or 'radam'.

  • learning_rate – Base learning rate for the optimizer.

  • sync_dist – Whether to sync logged metrics across devices.

class chronocratic.models.recurrent.enums.OptimizerName(*values)#

Bases: StrEnum

Optimizer variants for recurrent autoencoder training.

class chronocratic.models.recurrent.enums.ReconstructionLoss(*values)#

Bases: StrEnum

Reconstruction loss functions for the autoencoder objective.

class chronocratic.models.recurrent.enums.RecurrentCellType(*values)#

Bases: StrEnum

Recurrent cell variants for the autoencoder backbone.

TimeNet#

GRU-based encoder-decoder with autoencoder pretraining. Uses BasicEncodingMixin for inference.

class chronocratic.models.recurrent.timenet.model.TimeNet(hidden_dims: int, depth: int, input_dims: int = 1, dropout_rate: float = 0.1, learning_rate: float = 0.001)#

Bases: LightningModule, BasicEncodingMixin

TimeNet Model.

This model was implemented based on the code available on this GitHub repo paudan/TimeNet under MIT License.

property encoder: Sequential#

Return the GRU encoder.

property decoder: Sequential#

Return the GRU decoder.

forward(x: Tensor) → Tensor#

Reconstruct x from the reversed encoder sequence.

training_step(batch: Tensor, _batch_idx: int) → Tensor#

Compute and log the training reconstruction loss.

validation_step(batch: Tensor, _batch_idx: int) → Tensor#

Compute and log the validation reconstruction loss.

configure_optimizers() → OptimizerLRScheduler#

Return the Adam optimizer used to train TimeNet.

Configuration for the TimeNet model.

Provides TimeNetModelParameters with settings for the GRU encoder / decoder pair used by the autoencoder pretraining objective.

class chronocratic.models.recurrent.timenet.config.TimeNetModelParameters(*, hidden_dims: int, depth: int, input_dims: int = 1, dropout_rate: float = 0.1, learning_rate: float = 0.001)#

Bases: object

Configuration for the TimeNet model.

Parameters:

  • hidden_dims – Number of hidden units in each GRU layer of the encoder and decoder.

  • depth – Number of stacked GRU layers in each of the encoder and decoder.

  • input_dims – Number of input features (channels) in the time series.

  • dropout_rate – Dropout probability inserted between successive GRU layers. 0 disables dropout.

  • learning_rate – Base learning rate for the Adam optimizer.