针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的...针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的,通过在BNN的量化函数中引入可训练的量化阈值,使权重和激活值逐渐从三值演化到二值或零,旨在使网络在训练期间自动识别不重要的结构。此外,一个剪枝率调节算法也被设计用于调控网络的剪枝率。训练后,全零滤波器和对应的输出通道可被直接裁剪而获得精简的BNN,无需微调。为证明提出方法的可行性和其提升BNN推理效率而不牺牲准确率的潜力,在CIFAR-10上进行实验:在CIFAR-10数据集上,ETB对VGG-Small模型进行了46.3%的剪枝,模型大小压缩至0.34 MByte,准确率为89.97%,并在ResNet-18模型上进行了30.01%的剪枝,模型大小压缩至1.33 MByte,准确率为90.79%。在准确率和参数量方面,对比一些现有的BNN剪枝方法,ETB具有一定的优势。展开更多
Aeolian sand sample from Tengger desert, located in the southern part of Inner Mongolia (China) was characterized for major elemental composition and mineralogy by EPMA, XRF and XRD methods. The objective of this rese...Aeolian sand sample from Tengger desert, located in the southern part of Inner Mongolia (China) was characterized for major elemental composition and mineralogy by EPMA, XRF and XRD methods. The objective of this research was to provide data which would be a guide to aid future beneficiation of this sand, especially for the economic exploitation of feldspar and quartz which have a wide range of applications in various industries like plastic, paint, ceramics and glass industries. The elemental analysis of the sample was carried out by X-ray fluorescence spectrometer and chemical analysis while the minerals present were identified by an X-ray diffraction analyzer. The sand was discovered to contain basically SiO2 (82.43%), Al2O3 (7.68%), Na2O + K2O (4.37%) and TiO2 and Fe2O3 as the main impurities. It was also discovered that grinding of the sand is required to enhance the liberation of the minerals and the separation methods recommended are magnetic separation and flotation. It was therefore concluded that aeolian sand is a suitable source of quartz and feldspar for use in the industry.展开更多
文摘针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的,通过在BNN的量化函数中引入可训练的量化阈值,使权重和激活值逐渐从三值演化到二值或零,旨在使网络在训练期间自动识别不重要的结构。此外,一个剪枝率调节算法也被设计用于调控网络的剪枝率。训练后,全零滤波器和对应的输出通道可被直接裁剪而获得精简的BNN,无需微调。为证明提出方法的可行性和其提升BNN推理效率而不牺牲准确率的潜力,在CIFAR-10上进行实验:在CIFAR-10数据集上,ETB对VGG-Small模型进行了46.3%的剪枝,模型大小压缩至0.34 MByte,准确率为89.97%,并在ResNet-18模型上进行了30.01%的剪枝,模型大小压缩至1.33 MByte,准确率为90.79%。在准确率和参数量方面,对比一些现有的BNN剪枝方法,ETB具有一定的优势。
文摘Aeolian sand sample from Tengger desert, located in the southern part of Inner Mongolia (China) was characterized for major elemental composition and mineralogy by EPMA, XRF and XRD methods. The objective of this research was to provide data which would be a guide to aid future beneficiation of this sand, especially for the economic exploitation of feldspar and quartz which have a wide range of applications in various industries like plastic, paint, ceramics and glass industries. The elemental analysis of the sample was carried out by X-ray fluorescence spectrometer and chemical analysis while the minerals present were identified by an X-ray diffraction analyzer. The sand was discovered to contain basically SiO2 (82.43%), Al2O3 (7.68%), Na2O + K2O (4.37%) and TiO2 and Fe2O3 as the main impurities. It was also discovered that grinding of the sand is required to enhance the liberation of the minerals and the separation methods recommended are magnetic separation and flotation. It was therefore concluded that aeolian sand is a suitable source of quartz and feldspar for use in the industry.
