UniTS is a unified time series model designed to handle diverse tasks like forecasting, classification, imputation, and anomaly detection with a single, shared architecture. It targets researchers and practitioners seeking a versatile foundation model for time series data, offering strong zero-shot and few-shot learning capabilities across multiple domains without task-specific modules.

How It Works

UniTS employs a novel unified backbone featuring sequence and variable attention mechanisms, coupled with a dynamic linear operator. This design allows a single model to process various time series tasks and datasets without task-specific adaptations. The model is trained jointly across multiple datasets, enabling parameter sharing and promoting generalization.

Quick Start & Requirements

• Install: pip install -r requirements.txt (requires PyTorch 2.0+)
• Data: Run bash download_data_all.sh
• Training: Scripts are provided for multi-task learning, supervised learning, few-shot transfer learning (finetuning and prompt tuning), and zero-shot learning across various tasks.
• Pretrained weights: Available in the checkpoints directory.
• Documentation: UniTS Project Page

Highlighted Details

• Achieves superior performance over task-specific models and repurposed LLMs on 38 multi-domain datasets.
• Demonstrates remarkable zero-shot, few-shot, and prompt learning capabilities on new domains and tasks.
• Supports classification, forecasting, imputation, and anomaly detection with a single shared model.

Maintenance & Community

The project is associated with authors from Harvard University and MIT. It is built upon the Time-Series-Library.

Licensing & Compatibility

The material is approved for public release with unlimited distribution. It is provided "As-Is" and delivered to the U.S. Government with Unlimited Rights. Use may be restricted by existing copyrights.

Limitations & Caveats

Some zero-shot learning scripts require a specially trained version of UniTS with shared prompt/mask tokens. The README implies that specific pretrained checkpoints are necessary for few-shot and zero-shot transfer learning scripts.