Neural spatiotemporal forecasting with PyTorch
Torch Spatiotemporal (tsl) is a python library for neural spatiotemporal data processing, with a focus on Graph Neural Networks.
It is built upon the most used libraries of the python scientific computing ecosystem, with the final objective of providing a straightforward process that goes from data preprocessing to model prototyping. In particular, tsl offers a wide range of utilities to develop neural networks in PyTorch and PyG for processing spatiotemporal data signals.
In detail, the package provide:
High-level and easy-to-use APIs to build you own datasets and models for sensor networks.
Tools to deal with irregularities in the data stream: missing data, variations in the underlying network, etc.
Automatization of the preprocessing phase, with methods to scale and detrend the time series (see Preprocessing section).
A set of most used datasets in spatiotemporal data processing literature (see Datasets section).
A straightforward way of building spatiotemporal datasets that work with PyTorch and PyG (see Data structures section).
Out-of-the-box scalability – from a single CPU to clusters of GPUs – with PyTorch Lightning (see Inference engines section).
Plug-and-play state-of-the-art models from neural spatiotemporal literature (see Models section).
A collection of neural layers for creating neural spatiotemporal models in a fast and modular way (see Layers section).
A standard for experiment reproducibility based on the Hydra framework, to promote and support research on spatiotemporal data mining (see Experiment section).
“If I have seen further it is by standing on the shoulders of Giants.”
tsl relies heavily on these libraries for its functionalities: