shredx.datasets.dataloaders.TimeSeriesDataset

class shredx.datasets.dataloaders.TimeSeriesDataset(input_tensors: list[Float[Tensor, 'time_steps ...']], length: int, output_tensors: list[Float[Tensor, 'time_steps ...']] | None = None, device: str = 'cpu')

Bases: Dataset

Dataset of sliding windows over time series tensors.

Builds a dataset from lists of input and output tensors, each of shape (time_steps, ...). Assumes input and output tensors have the same shape. Input and output windows are returned separately to support use with proper orthogonal decomposition; they can be split or concatenated as needed when used.

Parameters:

input_tensorslist of torch.Tensor

List of input tensors, each a time series of shape (time_steps, ...).

lengthint

Length of each sliding window.

output_tensorslist of torch.Tensor, optional

List of output tensors, each of shape (time_steps, ...). If None, output windows are the same as input windows. Default is None.

devicestr, optional

Device to place tensors on. Default is "cpu".

Methods

prepare_tensors(tensors)

Prepare a list of tensors for use in the dataset.

Raises:

ValueError

If no valid windows can be created (e.g. tensor lengths too short for length).

Notes

Class Methods:

__len__():

• Returns the total number of valid windows across all tensors.

• Returns:

  • int. Total number of windows.

__getitem__(index):

• Returns the input and output windows for a given index.

• Parameters:

  • index : int. Index of the window to return.

• Returns:

  • tuple of (Float[torch.Tensor, "length ..."], Float[torch.Tensor, "length ..."]). (input_window, output_window), each of shape (length, ...).

prepare_tensors(tensors):

• Converts a list of tensors to float32 and moves them to the configured device.

• Parameters:

  • tensors : list of Float[torch.Tensor, "time_steps ..."]. List of tensors to prepare.

• Returns:

  • list of Float[torch.Tensor, "time_steps ..."]. Tensors on the correct device in float32.