随着移动云计算和边缘计算的快速发展，以及人工智能的广泛应用，产生了边缘智能(Edge Intelligence)的概念。深度神经网络(例如CNN)已被广泛应用于移动智能应用程序中，但是移动设备有限的存储和计算资源无法满足深度神经网络计算的需求。神经网络压缩与加速技术可以加速神经网络的计算，例如剪枝、量化、卷积核分解等。但是这些技术在实际应用非常复杂，并且可能导致模型精度的下降。在移动云计算或边缘计算中，任务卸载技术可以突破移动终端的资源限制，减轻移动设备的计算负载并提高任务处理效率。通过任务卸载技术优化深度神经网络成为边缘智能研究中的新方向。Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge这篇文章提出了协同推断的思想，将深度神经网络进行分区，一部分层在移动端计算，而另一部分在云端计算。根据硬件平台、无线网络以及服务器负载等因素实现动态分区，降低时延以及能耗。本项目给出了边缘智能方面的相关论文，并且给出了一个Python语言实现的卷积神经网络协同推断实验平台。关键词：边缘智能(Edge Intelligence)，计算卸载(Computing Offloading)，CNN模型分区(CNN Partition)，协同推断(Collaborative Inference)，移动云计算(Mobile Cloud Computing)
279
stars
107
forks
source link
对Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge论文中某些地方不理解，请问大佬能解答一下吗 #11
5.1 Performance Prediction Model
Neurosurgeon models the per-layer latency and the energy consumption of arbitrary neural network architecture. This approach allows Neurosurgeon to estimate the latency and energy consumption of a DNN's constituent layers without executing the DNN.
We observe that for each layer type, there is a large latency variation across layer configurations. Thus, to construct the prediction model for each layer type, we vary the configurable parameters of the layer and measure the latency and power consumption for each configuration. Using these profiles, we establish a regression model for each layer type to predict the latency and power of the layer based on its configuration. We describe each layer's regression model variables later in this section. We use GFLOPS (Giga Floating Point Operations per Second) as our performance metric. Based on the layer type, we use either a logarithmic or linear function as the regression function. The logarithmic-based regression is used to model the performance plateau as the computation requirement of the layer approaches the limit of the available hardware resources.
wyc941012
commented
5 months ago
在 5.1 Performance Prediction Model中，提到的GFLOPS起到什么作用？模型的输入时各种层的配置参数，输出是层的执行时间吗？
5.1 Performance Prediction Model Neurosurgeon models the per-layer latency and the energy consumption of arbitrary neural network architecture. This approach allows Neurosurgeon to estimate the latency and energy consumption of a DNN's constituent layers without executing the DNN.
We observe that for each layer type, there is a large latency variation across layer configurations. Thus, to construct the prediction model for each layer type, we vary the configurable parameters of the layer and measure the latency and power consumption for each configuration. Using these profiles, we establish a regression model for each layer type to predict the latency and power of the layer based on its configuration. We describe each layer's regression model variables later in this section. We use GFLOPS (Giga Floating Point Operations per Second) as our performance metric. Based on the layer type, we use either a logarithmic or linear function as the regression function. The logarithmic-based regression is used to model the performance plateau as the computation requirement of the layer approaches the limit of the available hardware resources.
最后一句当层的计算需求接近可用硬件资源的极限时，使用基于对数的回归对性能平台进行建模，我不太理解，大佬能解答一下吗