LB-PTQ: Effective Low-Bit Post-Training Quantization for Vision Transformers.

Abstract

Recently, Vision Transformers (ViTs) have become the state-of-the-art architecture on various computer vision tasks including image classification, object detection and semantic segmentation. However, such success in high-accuracy performance comes at the price of high computational complexity, with typically tens of millions of or even more parameters in a Vision Transformer (ViT) model. Such a large volume of parameters makes it very difficult to deploy ViT models on mobile devices and cumbers their applications. In this paper, we present a novel post-training quantization approach that is able to quantize ViT models to very low bit widths, without the need of re-training. Prior works on post-training quantization for ViTs optimize the quantization of each layer separately thus leading to sub-optimal results. In contrast, we propose a unified learning framework that jointly optimizes the quantization of all layers to directly reduce the overall output error of the network. Moreover, we explore an important property of ViTs, i.e., the additivity property, revealing that the output error caused by the quantization of multiple layers equals the sum of the output error due to the quantization of each layer. Utilizing this property, we present a very efficient algorithm to solve the joint optimization problem with linear time complexity. We performed extensive experiments on the large-scale ImageNet dataset to evaluate the effectiveness of our approach. Empirical results show that our approach improves state-of-the-art noticeably on various ViT models and lowers the bit width from 8-bit to 6-bit without hurting the accuracy. Specifically, at 4 bits, our approach significantly outperforms existing works by 1.72%, 11.49%, 6.15%, and 3.54% on ViT-S, ViT-B, DeiT-S, and DeiT-B, respectively. In the end, we evaluate the performance when deploying our quantized models on hardware. Our approach achieves $1.5\times $ to $1.7\times $ speedups for the inference on NVIDIA A100 GPU.