Differentiable Dynamic Quantization with Mixed Precision and Adaptive Resolution

TL;DR

This paper introduces Differentiable Dynamic Quantization (DDQ), a fully learnable method for quantizing neural networks that adapts precision and resolution per layer, outperforming prior methods especially on lightweight models like MobileNets.

Contribution

The paper proposes DDQ, a novel differentiable approach to learn quantization hyperparameters, enabling efficient and hardware-friendly low-precision neural network deployment.

Findings

DDQ achieves lossless 4-bit quantization for MobileNetV2 on ImageNet.

DDQ outperforms prior quantization methods on various networks and benchmarks.

DDQ is capable of quantizing challenging lightweight architectures effectively.

Abstract

Model quantization is challenging due to many tedious hyper-parameters such as precision (bitwidth), dynamic range (minimum and maximum discrete values) and stepsize (interval between discrete values). Unlike prior arts that carefully tune these values, we present a fully differentiable approach to learn all of them, named Differentiable Dynamic Quantization (DDQ), which has several benefits. (1) DDQ is able to quantize challenging lightweight architectures like MobileNets, where different layers prefer different quantization parameters. (2) DDQ is hardware-friendly and can be easily implemented using low-precision matrix-vector multiplication, making it capable in many hardware such as ARM. (3) Extensive experiments show that DDQ outperforms prior arts on many networks and benchmarks, especially when models are already efficient and compact. e.g., DDQ is the first approach that achieves lossless 4-bit quantization for MobileNetV2 on ImageNet.