// the find
zhouhaoyi/Informer2020
The GitHub repository for the paper "Informer" accepted by AAAI 2021.
Reference implementation for the AAAI'21 Best Paper that introduced ProbSparse attention — a way to make Transformers practical for long-horizon time series forecasting by approximating full attention in O(L log L) instead of O(L²). This is a research artifact, not a library. If you're doing academic work on time series transformers, this is the canonical starting point everyone else cites.
ProbSparse attention is a real algorithmic contribution: sampling the top-u queries by KL divergence from the uniform distribution is clever and well-motivated in the paper. The distilling operation (halving the sequence length between encoder layers) compounds the efficiency gain in a way that's clean to follow in the code. Docker + Makefile setup means you can actually reproduce the paper's numbers without archaeology. The argument table in the README is thorough — you won't be guessing what half the CLI flags do.
Pinned to Python 3.6 and PyTorch 1.2/1.8 — these are five-plus years old; you'll spend time fighting dependency conflicts before you write a line of your own code. The promised 'Informer V2' announcement (March 2023) never shipped, which is a bad sign for continued development. Subsequent papers (PatchTST, iTransformer, TimesNet) have largely outperformed Informer on the same ETT benchmarks it was evaluated on, so adopting this architecture for production forecasting now is questionable. License is CC BY-NC-SA, which means no commercial use — easy to miss.