UPDATE: Zenodo migration postponed to Oct 13 from 06:00-08:00 UTC. Read the announcement.
There is a newer version of this record available.

Software Open Access

torchtime: Time series data sets for PyTorch

Darke, Philip


  • Processed data are now cached in the .torchtime directory
  • train_split and val_split arguments are renamed train_prop and val_prop respectively
  • Introduced generic torchtime.data_TimeSeriesDataSet class behind the scenes - note training/validation/test data splits have changed for a given seed
  • torchtime.collate.packed_sequence now returns both X and y as a PackedSequence object
  • Expanded unit tests - note coverage is currently limited as PhysioNet2019 tests cannot be run under CI
  • Updated documentation


  • impute argument to support missing data imputation using mean and forward imputation methods or a custom imputation function
  • downscale argument to reduce the size of data sets for testing/model development
  • torchtime.data.TensorTimeSeriesDataset class to create a data set from input tensors


  • Use float32/torch.float and int64/torch.long precision for all data sets
  • Shape of y data in PhysioNet2019 data
  • Bug when adding time delta channels without a missing data mask
If you use this software, please cite it as below.
Files (36.6 kB)
Name Size
36.6 kB Download
All versions This version
Views 13743
Downloads 52
Data volume 216.6 kB73.2 kB
Unique views 11238
Unique downloads 52


Cite as