GitHub - sktime/pytorch-forecasting: Time series forecasting with PyTorch

Abstraction: High-level PyTorch library for deep learning time series forecasting

Key points:

• PyTorch Forecasting provides a high-level API built on PyTorch Lightning for training state-of-the-art time series models on GPU/CPU with automatic logging
• Includes Temporal Fusion Transformer (outperforms DeepAR by 36-69% on benchmarks), N-BEATS (won M4 competition as ensemble), N-HiTS (beats N-BEATS, good for long-horizon), and DeepAR
• TimeSeriesDataSet class abstracts variable transformations, missing values, randomized subsampling, and multiple history lengths
• Supports multi-horizon metrics, hyperparameter tuning via Optuna, and in-built interpretation capabilities
• Install via pip install pytorch-forecasting; optional MQF2 multivariate quantile loss support
• Models are configured primarily from the dataset metadata, minimizing manual hyperparameter specification

Connections: Pytorch · Pytorch Forecasting · Time Series Forecasting · Temporal Fusion Transformer · Deep Learning

Source: https://github.com/jdb78/pytorch-forecasting