人工智能培训

搜索

深度学习论文:DNQ:动态网络量化(DNQ: Dynamic Network Quantization)

[复制链接]
sa_group 发表于 2018-12-7 11:37:05 | 显示全部楼层 |阅读模式
sa_group 2018-12-7 11:37:05 32 0 显示全部楼层
深度学习论文:DNQ:动态网络量化(DNQ: Dynamic Network Quantization)网络量化是在存储器和能量受限的移动设备上部署神经网络的有效方法。在本文中,我们提出了一个动态网络量化(DNQ)框架,它由两个模块组成:一个位宽控制器和一个量化器。与大多数使用通用量化比特宽度的现有量化方法不同,我们利用策略梯度来训练代理以通过比特宽度控制器来学习每个层的比特宽度。该控制器可以在精度和压缩比之间进行权衡。给定量化比特宽度序列,量化器采用量化距离作为量化期间权重重要性的标准。我们在各种主流神经网络上广泛验证了所提出的方法,并获得了令人印象深刻的结果。
Network quantization is an effective method for the deployment of neuralnetworks on memory and energy constrained mobile devices.In this paper, wepropose a Dynamic Network Quantization (DNQ) framework which is composed of twomodules: a bit-width controller and a quantizer.Unlike most existingquantization methods that use a universal quantization bit-width for the wholenetwork, we utilize policy gradient to train an agent to learn the bit-width ofeach layer by the bit-width controller.This controller can make a trade-offbetween accuracy and compression ratio.Given the quantization bit-widthsequence, the quantizer adopts the quantization distance as the criterion ofthe weights importance during quantization.We extensively validate theproposed approach on various main-stream neural networks and obtain impressiveresults.深度学习论文:DNQ:动态网络量化(DNQ: Dynamic Network Quantization) TTLKbJzA3JvCOM9O.jpg
URL地址:https://arxiv.org/abs/1812.02375     ----pdf下载地址:https://arxiv.org/pdf/1812.02375    ----深度学习论文:DNQ:动态网络量化(DNQ: Dynamic Network Quantization)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

sa_group当前离线
新手上路

查看:32 | 回复:0

快速回复 返回顶部 返回列表