tsai

tsai is a Python library for researchers and developers building machine learning models for time series and sequences. Built on PyTorch and fastai, it brings model training, prediction and data preparation into one package. It's open source under Apache 2.0.

Its scope includes classification, regression, forecasting and filling missing values. Classification covers both single-variable and multivariable sequences, with binary or multiple-class targets. Forecasts can predict one or several steps ahead, using multiple input variables and producing multiple output variables.

The model selection is a reason to consider it for comparing approaches. PatchTST and TST provide transformer models, while InceptionTime and TCN use convolutional architectures. LSTM and GRU models include attention variants. Rocket and MiniRocket offer alternatives that use convolutions without being deep learning models; their integration uses sktime. The library also includes tabular models such as TabTransformer and TabFusionTransformer.

Data preparation tools include standardization, sliding windows and scikit-learn-style pipeline transforms. Walk-forward cross-validation supports evaluation over successive time periods. Trained models can be saved and loaded for prediction on other data.

tsai works in Python notebooks, and its tutorials include Google Colab examples. It provides downloadable datasets for classification, regression and forecasting, including long-term forecasting tasks.