torchtime 0.2.0

Time series data sets for PyTorch

Time series data sets for PyTorch

Ready-to-go PyTorch data sets for supervised time series prediction problems. torchtime currently supports:

• All data sets in the UEA/UCR classification repository [link]

• PhysioNet Challenge 2019 (sepsis prediction) [link]

Installation

$ pip install torchtime

Example usage

torchtime.data contains a class for each data set above. Each class has a consistent API.

The torchtime.data.UEA class returns the UEA/UCR data set specified by the dataset argument (see list of data sets here). For example, to load training data for the ArrowHead data set with a 70/30% training/validation split and create a DataLoader:

from torch.utils.data import DataLoader
from torchtime.data import UEA

arrowhead = UEA(
    dataset="ArrowHead",
    split="train",
    train_prop=0.7,
    seed=123
)
dataloader = DataLoader(arrowhead, batch_size=32)

Batches are dictionaries of tensors X, y and length.

X are the time series data. The package follows the batch first convention therefore X has shape (n, s, c) where n is batch size, s is (maximum) trajectory length and c is the number of channels. By default, a time stamp is appended to the time series data as the first channel.

y are one-hot encoded labels of shape (n, l) where l is the number of classes and length are the length of each trajectory (before padding if series are of irregular length) i.e. a tensor of shape (n).

ArrowHead is a univariate time series therefore X has two channels, the time stamp followed by the time series (c = 2). Each series has 251 observations (s = 251) and there are three classes (l = 3).

next_batch = next(iter(dataloader))

next_batch["X"].shape  # (32, 251, 2)
next_batch["y"].shape  # (32, 3)
next_batch["length"].shape  # (32)

Additional options

• The split argument determines whether training, validation or test data are returned. The size of the splits are controlled with the train_prop and val_prop arguments.

• Missing data can be imputed by setting impute to mean (replace with training data channel means) or forward (replace with previous observation). Alternatively a custom imputation function can be used.

• A time stamp, missing data mask and the time since previous observation can be appended to the time series data with the boolean arguments time, mask and delta respectively.

• For reproducibility, an optional random seed can be specified.

Most UEA/UCR data sets are regularly sampled and fully observed. Missing data can be simulated using the missing argument to drop data at random from UEA/UCR data sets. See the tutorials and API for more information.

Acknowledgements

torchtime uses some of the data processing ideas in Kidger et al, 2020 [1] and Che et al, 2018 [2].

This work is supported by the Engineering and Physical Sciences Research Council, Centre for Doctoral Training in Cloud Computing for Big Data, Newcastle University (grant number EP/L015358/1).

References

1. Kidger, P, Morrill, J, Foster, J, et al. Neural Controlled Differential Equations for Irregular Time Series. arXiv 2005.08926 (2020). [arXiv]

2. Che, Z, Purushotham, S, Cho, K, et al. Recurrent Neural Networks for Multivariate Time Series with Missing Values. Sci Rep 8, 6085 (2018). [doi]

3. Reyna, M, Josef, C, Jeter, R, et al. Early Prediction of Sepsis From Clinical Data: The PhysioNet/Computing in Cardiology Challenge. Critical Care Medicine 48 2: 210-217 (2019). [doi]

4. Reyna, M, Josef, C, Jeter, R, et al. Early Prediction of Sepsis from Clinical Data: The PhysioNet/Computing in Cardiology Challenge 2019 (version 1.0.0). PhysioNet (2019). [doi]

5. Goldberger, A, Amaral, L, Glass, L, et al. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation 101 (23), pp. e215–e220 (2000). [doi]

6. Löning, M, Bagnall, A, Ganesh, S, et al. sktime: A Unified Interface for Machine Learning with Time Series. Workshop on Systems for ML at NeurIPS 2019 (2019). [doi]

7. Löning, M, Bagnall, A, Middlehurst, M, et al. alan-turing-institute/sktime: v0.10.1 (v0.10.1). Zenodo (2022). [doi]

License

Released under the MIT license.