Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Deng, Zihao, Sharify, Sayeh, Wang, Xin, Orshansky, Michael
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2307.05657
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910905835454464
author Deng, Zihao
Sharify, Sayeh
Wang, Xin
Orshansky, Michael
author_facet Deng, Zihao
Sharify, Sayeh
Wang, Xin
Orshansky, Michael
contents Quantization is a widely used technique to compress neural networks. Assigning uniform bit-widths across all layers can result in significant accuracy degradation at low precision and inefficiency at high precision. Mixed-precision quantization (MPQ) addresses this by assigning varied bit-widths to layers, optimizing the accuracy-efficiency trade-off. Existing sensitivity-based methods for MPQ assume that quantization errors across layers are independent, which leads to suboptimal choices. We introduce CLADO, a practical sensitivity-based MPQ algorithm that captures cross-layer dependency of quantization error. CLADO approximates pairwise cross-layer errors using linear equations on a small data subset. Layerwise bit-widths are assigned by optimizing a new MPQ formulation based on cross-layer quantization errors using an Integer Quadratic Program. Experiments with CNN and vision transformer models on ImageNet demonstrate that CLADO achieves state-of-the-art mixed-precision quantization performance. Code repository available here: https://github.com/JamesTuna/CLADO_MPQ
format Preprint
id arxiv_https___arxiv_org_abs_2307_05657
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Mixed-Precision Quantization for Deep Vision Models with Integer Quadratic Programming
Deng, Zihao
Sharify, Sayeh
Wang, Xin
Orshansky, Michael
Neural and Evolutionary Computing
Quantization is a widely used technique to compress neural networks. Assigning uniform bit-widths across all layers can result in significant accuracy degradation at low precision and inefficiency at high precision. Mixed-precision quantization (MPQ) addresses this by assigning varied bit-widths to layers, optimizing the accuracy-efficiency trade-off. Existing sensitivity-based methods for MPQ assume that quantization errors across layers are independent, which leads to suboptimal choices. We introduce CLADO, a practical sensitivity-based MPQ algorithm that captures cross-layer dependency of quantization error. CLADO approximates pairwise cross-layer errors using linear equations on a small data subset. Layerwise bit-widths are assigned by optimizing a new MPQ formulation based on cross-layer quantization errors using an Integer Quadratic Program. Experiments with CNN and vision transformer models on ImageNet demonstrate that CLADO achieves state-of-the-art mixed-precision quantization performance. Code repository available here: https://github.com/JamesTuna/CLADO_MPQ
title Mixed-Precision Quantization for Deep Vision Models with Integer Quadratic Programming
topic Neural and Evolutionary Computing
url https://arxiv.org/abs/2307.05657