Sub-byte quantization of Mobile Face Recognition Convolutional Neural Networks

Abstract

Converting convolutional neural networks such as MobileNets to a full integer representation is already quite a popular method to reduce the size and computational footprint of classification networks but its effect on face recognition networks is relatively unexplored. This work presents a method to reduce the size of MobileFaceNet using sub-byte quantization of the weights and activations. It was found that 8-bit and 4-bit versions of MobileFaceNet can be obtained with 98.68% and 98.63% accuracy on the LFW dataset which reduces the footprint to 25% and 12.5% of the original weights respectively. Using mixed-precision, an accuracy of 98.17% can be achieved whilst requiring only 10% of the original weight footprint. It is expected that with a larger training dataset, higher accuracies can be achieved.