Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference

TL;DR

This paper introduces an integer-only quantization scheme and a co-designed training method that enable efficient, accurate deep learning inference on mobile devices with integer hardware, improving latency and accuracy tradeoffs.

Contribution

It presents a novel quantization scheme and training procedure that together enable end-to-end integer-only inference with minimal accuracy loss.

Findings

Significant accuracy improvements on MobileNets after quantization.

Enhanced on-device inference speed and efficiency.

Validated on ImageNet and COCO datasets with CPU hardware.

Abstract

The rising popularity of intelligent mobile devices and the daunting computational cost of deep learning-based models call for efficient and accurate on-device inference schemes. We propose a quantization scheme that allows inference to be carried out using integer-only arithmetic, which can be implemented more efficiently than floating point inference on commonly available integer-only hardware. We also co-design a training procedure to preserve end-to-end model accuracy post quantization. As a result, the proposed quantization scheme improves the tradeoff between accuracy and on-device latency. The improvements are significant even on MobileNets, a model family known for run-time efficiency, and are demonstrated in ImageNet classification and COCO detection on popular CPUs.