MinGRU Implementation in Keras

This repository contains a Keras implementation of the minGRU model, a minimal and parallelizable version of the traditional Gated Recurrent Unit (GRU) architecture. The minGRU model is based on the research paper "Were RNNs All We Needed?" that revisits traditional recurrent neural networks and modifies them to be efficiently trained in parallel.

Features

• Minimal GRU architecture with significantly fewer parameters than traditional GRUs.
• Fully parallelizable during training, achieving faster training times.
• Compatible with Keras 3.

Installation

This project uses uv to manage dependencies. To install the required dependencies, run:

pip install --upgrade mingru-keras

Usage

To use the MinGRU model in your own project, simply import the MinGRU class and use it as you would any other Keras layer.

Example

>>> import keras
>>> from mingru_keras import MinGRU
>>> layer = MinGRU(units=64)
>>> X = keras.random.normal((32, 1000, 8))
>>> layer(X).shape
(32, 1000, 64)

Contributing

Contributions are welcome! If you'd like to report a bug or suggest a feature, please open an issue or submit a pull request.