ID 原文 译文
58728 并基于批归一化层的计算访存特征,将其重构为两个子层,分别与相邻层融合,进一步减少训练时对主存的读写,并构建了训练时的访存量模型与计算量模型。 Besides, the BN layer is split into two sub-layers which are respectively fused with its adjacent layers, and this approach further reduces memory access during training and effectively improves the accelerator’s computational efficiency.
58729 实验结果表明,使用NVIDIA TESLA V100 GPU训练ResNet-50、Inception V3及DenseNet模型时,同原始训练方法相比,其访存数据量分别降低了33%,22%及31%,V100的实际计算效率分别提升了20.5%,18. 5%以及18. 1%。 Experimental results show that the amount of memory access is decreased by 33%, 22% and 31% respectively, and the actual computing efficiency of the V100 is improved by 20. 5%, 18. 5% and 18. 1% respectively when the ResNet-50, Inception V3 and DenseNet are trained on the NVIDIA TELSA V100 GPU with the optimization method.
58730 这种优化方法利用了网络结构与模型训练时的访存特点,可与其他访存优化方法协同使用,进一步降低模型训练时的访存量。 The proposed method exploits the characteristics of memory access during training, and can be used in conjunction with other optimization methods to further reduce the amount of memory access during training.
58731 由于卫星影像的定位精度主要受卫星影像有理多项式参数的估计精度影响,而现有算法中通常采用像点补偿或利用地面控制点改正的方法,未考虑设计矩阵元素的误差特性,存在系统误差剔除不完全、参数估计精度不高等问题。 The positioning accuracy of a satellite image is mainly affected by the estimation accuracy of the rational polynomial coefficients (RPCs). Image point compensation or ground control point correction methods are usually used in the existing algorithms. Because the error characteristics of the design matrix elements are not considered, there are problems such as incomplete systematic error elimination and low parameter estimation accuracy.
58732 考虑到模型系统误差的影响,提出了一种异方差估计方法。 Considering the influence of the model systematic error, a heteroscedastic estimation method is proposed in this paper.
58733 首先建立了设计矩阵元素的随机模型,对系统特性进行更精准的描述; First, the random model of matrix elements is established in the algorithm to describe the system characteristics more accurately.
58734 然后将设计矩阵元素的系统偏差考虑在内,采用马氏距离构建最小二乘模型,并使用广义特征值方法进行参数求解,从理论上降低了模型的系统误差。 Taking into account the system deviations of the design matrix elements, the least square model is constructed using the Mahalanobis distance as the metric, and parameters are solved using the generalized eigenvalue method. The systematic error can be reduced theoretically.
58735 在天绘一号不同地形影像上的实验结果表明,新算法在影像纠正精度上较传统方法提升36倍以上,且精度一致性优良,对提高有理多项式参数估计和卫星影像定位精度具有重要意义。 Experiment on different terrain images of TH-1 shows that the image correction accuracy of the proposed method is improved by more than 36 times compared with the traditional method, and the precision consistency is superior, which is of great significance to improving the accuracy of RPC parameters estimation and satellite imagery positioning.
58736 为解决传统监督学习方法在不均衡数据集异常检测上易出现过拟合的问题,提出一种非监督对抗学习方法用于硬盘故障预测。 In order to solve the problem of over-fitting of traditional supervised learning methods in anomaly detection of unbalanced datasets, an unsupervised adversarial learning method is proposed for hard disk failure prediction.
58737 该方法使用长短期记忆神经网络和全连接层设计了一种可进行二次编码的自编码器,只需使用正常样本进行训练,通过减小样本重构误差和潜在向量之间的距离,使模型学习正常样本的数据分布,从而提高了模型的泛化能力。 This method uses the long short-term memory neural network and fully connected layer to design an Autoencoder that can be used for secondary coding. Only normal samples are used for training. By reducing the reconstruction error and the distance between potential vectors, the model can learn the data distribution of normal samples, thus improving the generalization ability of the model.