Quantisation Aware Training of Neural Networks

This demonstrator implements some QAT training techniques. Currently the following techniques are implemented either natively or though Larq

• BinaryConnect: Just run a normal network with quantizers on
• BinaryNet: A BinaryConnect with binary activations
• Bop
• A custom variant of Bop, Greedy

Logic

The library is based off Larq and thus inherits its logic. We recommend you head to its documentation if you want to understand the underpinning logic. You can also directly go to their Larq github. Here is a brief synthesis:

• Everything is built off tensorflow
• Layers are replaced with quantized layers. Quantized layers take as a quantizer as a parameters, amongst other things. The quantizer dictates how the floating variable is transformed to a quantized variable, and how the gradient is backpropagated through the quantization step. This is what allows for QAT training.
• Optimizers are replaced by an overarching Case Optimizer. A Case Optimizer is a list of conditions and classic tensorflow optimizers. Whenever a variable satisfied the condition associated to an optimizer, that optimizer is used to update that variable. This allows applying specific optimizers to quantized variables

Documentation

Benchmarks

1. Methods and Tools for the Optimization of Machine Learning Components Experiments

Applications over use-cases

1. Impact of Quantization on Neural Network Stability Application on AirLiquide Cylinder Counting

Scientific contribution

1. Layer Level Sensitivity Assessment for Efficient and Robust Neural Network Inference

State of the Art

1. Methods and Tools for the Optimization of Machine Learning Components SOTA

Demonstrators

Restricted demonstrator is available here. it relies on a demand forecasting and a 2D scene understanding use-cases.

Support

Support for Quantization Aware Training must be obtained by sending an email to support@confiance.ai

Ensure your email contains :

• Your name
• A link to this page
• the version you are working with
• A clear description of the problematic (bug, crash, feature or help request)
• A full description of the problem whichallow to reproduce it
• Any file or screenshort element mandatory for the full understanding of the problematic