`Fusionlab-learn`: Igniting Next-Gen Fusion Models¶

A Modular Library for Temporal Fusion Transformer (TFT) Variants & Beyond

Extend, experiment, and fuse time-series predictions with state-of-the-art architectures.

Fusionlab-learn provides a flexible and extensible framework built on TensorFlow/Keras for advanced time-series forecasting. It centers on the Temporal Fusion Transformer (TFT) and its extensions like the Extreme Temporal Fusion Transformer (XTFT), offering modular components and powerful utilities for researchers and practitioners.

Whether you need interpretable multi-horizon forecasts, robust uncertainty quantification, or a platform to experiment with novel temporal architectures, FusionLab aims to provide the necessary tools.

Install fusion-lab

Quickstart

User Guide

API Reference

What’s New?

Release Notes

Key Features

🧩 Modular Design: Build custom forecasting models using interchangeable components like specialized attention layers, GRNs, VSNs, multi-scale LSTMs, and more. Facilitates research and tailored solutions.
🚀 Advanced Architectures: Includes robust implementations of standard TFT, the high-capacity XTFT for complex scenarios, and experimental SuperXTFT. Ready-to-use state-of-the-art models.
💡 Extensible: Designed for extension. Easily integrate new model architectures, custom layers, or novel loss functions to push the boundaries of time series forecasting.
⚙️ TensorFlow Backend: Currently leverages the power and scalability of the TensorFlow/Keras ecosystem for building and training models.
🛠️ Comprehensive Utilities: Offers a suite of helper tools for common tasks: data preparation, sequence generation, time series analysis, result visualization, hyperparameter tuning, and CLI applications.
🔬 Anomaly Detection: Features integrated anomaly detection mechanisms within XTFT, allowing models to identify and potentially adapt to irregular data patterns during training.

Explore Further

Motivation: Understand the Motivation behind FusionLab.
Examples: See practical applications in the Examples Gallery.
Contribute: Learn how to contribute to the project.
Cite: Find out how to cite FusionLab in your work.
Reference: Consult the Glossary or view the License (BSD-3-Clause).

Terminology

For brevity and consistency, the library will be referred to as fusionlab throughout the remainder of the documentation.

Fusionlab-learn: Igniting Next-Gen Fusion Models¶

`Fusionlab-learn`: Igniting Next-Gen Fusion Models¶