基于卷积神经网络的深度学习算法展现出卓越性能的同时也带来了冗杂的数据量和计算量,大量的存储与计算开销也成了该类算法在硬件平台部署过程中的最大阻碍。而神经网络模型量化使用低精度定点数代替原始模型中的高精度浮点数,在损失较...基于卷积神经网络的深度学习算法展现出卓越性能的同时也带来了冗杂的数据量和计算量,大量的存储与计算开销也成了该类算法在硬件平台部署过程中的最大阻碍。而神经网络模型量化使用低精度定点数代替原始模型中的高精度浮点数,在损失较小精度的前提下可有效压缩模型大小,减少硬件资源开销,提高模型推理速度。现有的量化方法大多将模型各层数据量化至相同精度,混合精度量化则根据不同层的数据分布设置不同的量化精度,旨在相同压缩比下达到更高的模型准确率,但寻找合适的混合精度量化策略仍十分困难。因此,提出一种基于误差限制的混合精度量化策略,通过对神经网络卷积层中的放缩因子进行统一等比限制,确定各层的量化精度,并使用截断方法线性量化权重和激活至低精度定点数,在相同压缩比下,相比统一精度量化方法有更高的准确率。其次,将卷积神经网络的经典目标检测算法YOLOV5s作为基准模型,测试了方法的效果。在COCO数据集和VOC数据集上,该方法与统一精度量化相比,压缩到5位的模型平均精度均值(mean Average Precision,mAP)分别提高了6%和24.9%。展开更多
The development of sustainable energy storage technologies is critical in addressing the global challenges posed by climate change.Supercapacitors,while offering exceptional power density and cycling stability,suffer ...The development of sustainable energy storage technologies is critical in addressing the global challenges posed by climate change.Supercapacitors,while offering exceptional power density and cycling stability,suffer from relatively low energy density,limiting their widespread use in large-scale energy storage systems.To overcome this limitation,we designed a novel composite electrode material featuring a core–shell structure.The core derived from well-defined ZIF-67 nanocubes(NCs)was innovatively processed into a hollow structure,which enhanced ion diffusion and increased the overall energy storage capacity by reducing internal resistance.Meanwhile,the shell consisted of 3D hierarchical Ni–Co layered double hydroxides(NiCo-LDH)grown in situ employing an ambient-temperature method,offering high electrochemical activity and abundant active sites for efficient charge storage.The ultimately synthesized multi-scale hollow core–shell material,Co_(3)O_(4)-HNC@NiCo-LDH,integrated the respective merits of the shell and core materials,while simultaneously addressing issues that arise when these materials exist in isolation.It effectively mitigated problems such as volume expansion and agglomeration that materials might encounter during electrochemical reactions,thereby further enhancing the materials’performance and service life.Notably,in situ Raman spectroscopy was utilized to trace the dynamic redox processes and structural changes occurring during electrochemical cycling,thereby validating the stability and effectiveness of the charge storage mechanism.The resulting material,Co_(3)O_(4)-HNC@NiCo-LDH,demonstrated impressive capacitance(1862.4 F g^(−1)at 2 A g^(−1)),high energy density(76.8 Wh kg^(−1)at 2 A g^(−1)),and excellent cycling stability(98.38%after 15000 cycles at 15 A g^(−1)),offering a promising solution for next-generation supercapacitors.展开更多
文摘基于卷积神经网络的深度学习算法展现出卓越性能的同时也带来了冗杂的数据量和计算量,大量的存储与计算开销也成了该类算法在硬件平台部署过程中的最大阻碍。而神经网络模型量化使用低精度定点数代替原始模型中的高精度浮点数,在损失较小精度的前提下可有效压缩模型大小,减少硬件资源开销,提高模型推理速度。现有的量化方法大多将模型各层数据量化至相同精度,混合精度量化则根据不同层的数据分布设置不同的量化精度,旨在相同压缩比下达到更高的模型准确率,但寻找合适的混合精度量化策略仍十分困难。因此,提出一种基于误差限制的混合精度量化策略,通过对神经网络卷积层中的放缩因子进行统一等比限制,确定各层的量化精度,并使用截断方法线性量化权重和激活至低精度定点数,在相同压缩比下,相比统一精度量化方法有更高的准确率。其次,将卷积神经网络的经典目标检测算法YOLOV5s作为基准模型,测试了方法的效果。在COCO数据集和VOC数据集上,该方法与统一精度量化相比,压缩到5位的模型平均精度均值(mean Average Precision,mAP)分别提高了6%和24.9%。
基金supported by the Science and Technology Development Plan Project of Jilin Province,China(No.20250102066JC)the Major Science and Technology Projects for Independent Innovation of China FAW Group Co.,Ltd(Grant No.20220301018GX and 20220301019GX).
文摘The development of sustainable energy storage technologies is critical in addressing the global challenges posed by climate change.Supercapacitors,while offering exceptional power density and cycling stability,suffer from relatively low energy density,limiting their widespread use in large-scale energy storage systems.To overcome this limitation,we designed a novel composite electrode material featuring a core–shell structure.The core derived from well-defined ZIF-67 nanocubes(NCs)was innovatively processed into a hollow structure,which enhanced ion diffusion and increased the overall energy storage capacity by reducing internal resistance.Meanwhile,the shell consisted of 3D hierarchical Ni–Co layered double hydroxides(NiCo-LDH)grown in situ employing an ambient-temperature method,offering high electrochemical activity and abundant active sites for efficient charge storage.The ultimately synthesized multi-scale hollow core–shell material,Co_(3)O_(4)-HNC@NiCo-LDH,integrated the respective merits of the shell and core materials,while simultaneously addressing issues that arise when these materials exist in isolation.It effectively mitigated problems such as volume expansion and agglomeration that materials might encounter during electrochemical reactions,thereby further enhancing the materials’performance and service life.Notably,in situ Raman spectroscopy was utilized to trace the dynamic redox processes and structural changes occurring during electrochemical cycling,thereby validating the stability and effectiveness of the charge storage mechanism.The resulting material,Co_(3)O_(4)-HNC@NiCo-LDH,demonstrated impressive capacitance(1862.4 F g^(−1)at 2 A g^(−1)),high energy density(76.8 Wh kg^(−1)at 2 A g^(−1)),and excellent cycling stability(98.38%after 15000 cycles at 15 A g^(−1)),offering a promising solution for next-generation supercapacitors.
