François Chollet, the creator of Keras, released version 3.1.0 of the library today. This version fixes key issues and improves upon model slicing and error handling,  it also extends support for more complex data handling techniques.

The update introduced features such as int8 quantisation for model optimisation, new backend configuration options, and a MelSpectrogram layer for audio data processing. 

Int8 quantisation, a technique that converts a model’s weights and activations from 32-bit floating-point to 8-bit integers, reduces the model’s size and speeds up its processes. This doesn’t cause a significant loss in accuracy but allows them to run more efficiently, especially on devices with limited resources.

Responding to the release a user on X said, “Quantisation was much needed! Hoping for QLoRA in the next version!”

New backend configuration options give developers flexibility to switch computational backends easily, enhancing the tool’s adaptability across different platforms like TensorFlow, or Microsoft Cognitive Toolkit (CNTK), depending on the developers preferences. 

In addition, the MelSpectrogram layer is added that transforms raw audio into Mel Spectrogram format that represents audio signals in a way that mimics human ear processing. This is particularly useful for machine learning models in tasks like speech recognition, audio classification, and music analysis. 

The update simplifies the development process, enhances model efficiency, and extends the library’s applicability across different computational environments and data types.