ibm-granite/granite-tsfm

What it solves

This project provides the tools and utilities necessary to work with Time Series Foundation Models (TSFM), specifically the Granite time series collection. It simplifies the process of using, pre-training, and fine-tuning these models for time series forecasting tasks.

How it works

The repository serves as a companion to the models hosted on Hugging Face. It provides a set of Python utilities, serving components, and a collection of notebooks that guide users through the process of pre-training and fine-tuning models like PatchTSMixer, PatchTST, and TinyTimeMixer (TTM).

Who it’s for

Data scientists and machine learning engineers who need to implement time series forecasting using foundation models and want a structured way to perform transfer learning or benchmarking.

Highlights

• Foundation Model Support: Supports a variety of models including TinyTimeMixer (TTM), PatchTSMixer, PatchTST, and FlowState.
• Integration with Hugging Face: Designed to work seamlessly with the models available in the Hugging Face transformers library.
• Practical Tutorials: Includes Google Colab tutorials and notebooks for getting started and performing transfer learning.
• Serving Components: Includes components for model inference and serving.