tsai

tsai is an open-source Python package from timeseriesAI for applying deep learning to time series and sequence data. It is built on PyTorch and fastai and covers tasks such as classification, regression, forecasting and imputation. The documentation site includes tutorial notebooks and API pages for data, models and training.

What It Is

tsai is a deep learning library for time series. Users prepare input as a 3-dimensional array (samples x variables x sequence length), or a pandas dataframe for tabular models, then train a model through learner classes such as TSClassifier, TSRegressor and TSForecaster. Trained learners can be exported and reloaded for inference.

Models and Data Tooling

The library ships many architectures, including LSTM, GRU, MLP, FCN, ResNet, InceptionTime, XceptionTime, TCN, ROCKET and MiniRocket, TST, TSiT, TSPerceiver and PatchTST. It also includes tabular models such as TabTransformer. Its data tooling covers downloadable benchmark datasets, sliding windows, time-based splitting, standardization, augmentation, and imaging or featurizing of time series. The docs also list sklearn-type pipeline transforms and walk-forward cross validation.

Forecasting and Training Workflow

For forecasting, the docs describe univariate or multivariate input and output, with single-step or multi-step horizons. PatchTST and models ending in Plus auto-configure their output head to match the target shape. Training uses fastai-style fit_one_cycle with callbacks, and the docs cover hyperparameter optimization with Optuna and Weights & Biases sweeps, as well as calibration, analysis and explainability.

Setup Path

Install with pip (Python 3.10 or newer is required as of tsai 1.0.0). Optional extras such as sktime are needed for the ROCKET models. Conda distribution is no longer updated. A Colab introduction to time series classification is the suggested starting point.