tsai: a deep learning library for time series and sequential data with a vast collection of SOTA models

What it solves

tsai is a deep learning library designed to simplify the implementation of state-of-the-art techniques for time series and sequential data. It provides a unified framework for tasks such as classification, regression, forecasting, and imputation, reducing the complexity of building and training these models.

How it works

Built on top of PyTorch and fastai, tsai provides a high-level API for training and inference. It supports a wide variety of model architectures, including LSTMs, GRUs, Transformers (like PatchTST and TST), and specialized time series models like InceptionTime and MiniRocket. The library handles data preparation through tools like SlidingWindow and TSStandardize, and expects input data in a 3D array format ([samples x variables x sequence length]) for time series models.

Who it’s for

Data scientists, ML engineers, and researchers who need to apply deep learning to time series data for forecasting or classification tasks.

Highlights

• Extensive Model Zoo: Includes a vast array of SOTA models ranging from traditional RNNs to modern Transformers and CNNs.
• Comprehensive Dataset Access: Built-in support for downloading over 200 univariate and multivariate datasets for classification, regression, and forecasting.
• Flexible Forecasting: Supports single-step and multi-step ahead forecasting for both univariate and multivariate inputs/outputs.
• Integrated Pipeline: Provides sklearn-type pipeline transforms and walk-forward cross-validation to improve model accuracy and evaluation.