// the find
zalandoresearch/pytorch-ts
PyTorch based Probabilistic Time Series forecasting framework based on GluonTS backend
PyTorchTS wraps GluonTS's data pipeline and backtesting infrastructure around PyTorch implementations of probabilistic forecasting models — DeepAR, N-BEATS, TFT, TimeGrad (diffusion-based), and a few others. It's aimed at researchers and practitioners who want GluonTS's evaluation harness but don't want MXNet anywhere near their stack. The original author is now at Hugging Face working on similar things, which explains the maintenance trajectory.
Includes TimeGrad (autoregressive diffusion) and TempFlow (normalizing flows) that were novel at publication time and aren't in every other forecasting library. The GluonTS backend gives you backtesting, dataset loaders, and probabilistic metrics without reimplementing them. Model implementations are clean enough to read as reference code — the separation between estimator, network, and output modules is consistent across all models. Probabilistic outputs are first-class: you get full predictive distributions, not just point estimates.
Last commit was June 2024 and the repo has the feel of research code that was never meant to be a maintained library — open issues go unanswered for years. GluonTS itself has since added PyTorch support natively, which makes this wrapper somewhat redundant; if you're starting fresh today, you'd likely just use GluonTS directly. The Trainer is a thin hand-rolled loop with no gradient accumulation, mixed precision, or distributed training support. No model checkpointing, no early stopping — you have to wire that yourself.