iTransformer

Why we included this project

This is a clean, self-contained implementation of the iTransformer architecture from Tsinghua and Ant Group, built for multivariate time series forecasting. Instead of treating each timestamp as a token, it flips the usual setup and treats each variate as a token, so attention models relationships across channels while a feed-forward network handles temporal dynamics. The package installs via pip and exposes a straightforward PyTorch API, letting you set lookback length, prediction horizons, and model depth in a few lines. It also includes experimental variants, like a 2D version that attends across both time and variates and an FFT-token variant, if you want to go beyond the paper's baseline. Teams working on energy, finance, or sensor data will find it a practical starting point before moving to the official reference implementation.