mct-quantizers-nightly 1.1.0.20230703.post1217

Infrastructure for support neural networks compression

Model Compression Toolkit (MCT) Quantizers

The MCT Quantizers library is an open-source library developed by researchers and engineers working at Sony Semiconductor Israel.

It provides tools for easily representing a quantized neural network in both Keras and PyTorch. The library offers researchers, developers, and engineers a set of useful quantizers, along with a simple interface for implementing new custom quantizers.

High level description:

The library's quantizers interface consists of two main components:

1. QuantizationWrapper: This object takes a layer with weights and a set of weight quantizers to infer a quantized layer.
2. ActivationQuantizationHolder: An object that holds an activation quantizer to be used during inference.

Users can set the quantizers and all the quantization information for each layer by initializing the weights_quantizer and activation_quantizer API.

Please note that the quantization wrapper and the quantizers are framework-specific.

Quantizers:

The library provides the "Inferable Quantizer" interface for implementing new quantizers. This interface is based on the BaseInferableQuantizer class, which allows the definition of quantizers used for emulating inference-time quantization.

On top of BaseInferableQuantizer the library defines a set of framework-specific quantizers for both weights and activations:

1. Keras Quantizers
2. Pytorch Quantizers

The mark_quantizer Decorator

The @mark_quantizer decorator is used to assign each quantizer with static properties that define its task compatibility. Each quantizer class should be decorated with this decorator, which defines the following properties:

• QuantizationTarget: An Enum that indicates whether the quantizer is intended for weights or activations quantization.
• QuantizationMethod: A list of quantization methods (Uniform, Symmetric, etc.).
• quantizer_type: An optional property that defines the type of the quantization technique. This is a helper property that allows the creation of advanced quantizers for specific tasks.

Getting Started

This section provides a quick guide to getting started. We begin with the installation process, either via source code or the pip server. Then, we provide a short example of usage.

Installation

Please refer to the MCT install guide for installing the pip package or building from the source.

From Source

git clone https://github.com/sony/mct_quantizers.git
python setup.py install

From PyPi - nightly package

Currently, only a nightly released package (unstable) is available via PyPi.

pip install mct-quantizers-nightly

Requirements

To use MCT Quantizers, you need to have one of the supported frameworks, Tensorflow or PyTorch, installed.

For use with Tensorflow, please install the following packages: tensorflow, tensorflow-model-optimization

For use with PyTorch, please install the following package: torch

You can also use the requirements file to set up your environment.

License

Apache License 2.0.