近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度...近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度类型,使网络体积相较于特定硬件平台上的内存空间更大,且floating point 16、INT 8等单精度类型已无法满足现在一些模型推理的现实需求。为此,提出一种以子图为最小单位,通过判断相邻结点之间的融合关系,添加了丰富比特位的混合精度推理算法。首先,在原有单精度量化设计的搜索空间中增加floating point 16半精度的比特配置,使最终搜索空间变大,为寻找最优解提供更多机会。其次,使用子图融合的思想,通过整数线性规划将融合后的不同子图精度配置,根据模型大小、推理延迟和位宽操作数3个约束对计算图进行划分,使最后累积的扰动误差减少。最终,在ResNet系列网络上验证发现,所提模型精度相较于HAWQ V3的损失没超过1%的同时,相较于其他混合精度量化方法在推理速度方面得到了提升,在ResNet18网络中推理速度分别提升18.15%、19.21%,在ResNet50网络中推理速度分别提升13.15%、13.70%。展开更多
Noise level in a marine environment has raised extensive concern in the scientific community.The research is carried out on i4 Ocean platform following the process of ocean noise model integrating,noise data extractin...Noise level in a marine environment has raised extensive concern in the scientific community.The research is carried out on i4 Ocean platform following the process of ocean noise model integrating,noise data extracting,processing,visualizing,and interpreting,ocean noise map constructing and publishing.For the convenience of numerical computation,based on the characteristics of ocean noise field,a hybrid model related to spatial locations is suggested in the propagation model.The normal mode method K/I model is used for far field and ray method CANARY model is used for near field.Visualizing marine ambient noise data is critical to understanding and predicting marine noise for relevant decision making.Marine noise map can be constructed on virtual ocean scene.The systematic marine noise visualization framework includes preprocessing,coordinate transformation interpolation,and rendering.The simulation of ocean noise depends on realistic surface.Then the dynamic water simulation gird was improved with GPU fusion to achieve seamless combination with the visualization result of ocean noise.At the same time,the profile and spherical visualization include space,and time dimensionality were also provided for the vertical field characteristics of ocean ambient noise.Finally,marine noise map can be published with grid pre-processing and multistage cache technology to better serve the public.展开更多
文摘近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度类型,使网络体积相较于特定硬件平台上的内存空间更大,且floating point 16、INT 8等单精度类型已无法满足现在一些模型推理的现实需求。为此,提出一种以子图为最小单位,通过判断相邻结点之间的融合关系,添加了丰富比特位的混合精度推理算法。首先,在原有单精度量化设计的搜索空间中增加floating point 16半精度的比特配置,使最终搜索空间变大,为寻找最优解提供更多机会。其次,使用子图融合的思想,通过整数线性规划将融合后的不同子图精度配置,根据模型大小、推理延迟和位宽操作数3个约束对计算图进行划分,使最后累积的扰动误差减少。最终,在ResNet系列网络上验证发现,所提模型精度相较于HAWQ V3的损失没超过1%的同时,相较于其他混合精度量化方法在推理速度方面得到了提升,在ResNet18网络中推理速度分别提升18.15%、19.21%,在ResNet50网络中推理速度分别提升13.15%、13.70%。
基金supported by the Global Change and Air-Sea Interaction Project (GASI-03-01-01-09)the National Natural Science Foundation of China under Grant No.61170106+1 种基金the Project funded by China Postdoctoral Science Foundation under Grant No.2015M571993Qingdao Postdoctoral Application Research Funded Project
文摘Noise level in a marine environment has raised extensive concern in the scientific community.The research is carried out on i4 Ocean platform following the process of ocean noise model integrating,noise data extracting,processing,visualizing,and interpreting,ocean noise map constructing and publishing.For the convenience of numerical computation,based on the characteristics of ocean noise field,a hybrid model related to spatial locations is suggested in the propagation model.The normal mode method K/I model is used for far field and ray method CANARY model is used for near field.Visualizing marine ambient noise data is critical to understanding and predicting marine noise for relevant decision making.Marine noise map can be constructed on virtual ocean scene.The systematic marine noise visualization framework includes preprocessing,coordinate transformation interpolation,and rendering.The simulation of ocean noise depends on realistic surface.Then the dynamic water simulation gird was improved with GPU fusion to achieve seamless combination with the visualization result of ocean noise.At the same time,the profile and spherical visualization include space,and time dimensionality were also provided for the vertical field characteristics of ocean ambient noise.Finally,marine noise map can be published with grid pre-processing and multistage cache technology to better serve the public.
