【作者】 张宇；

【导师】 邢永康；

【作者基本信息】 重庆大学 ， 工程（计算机技术）（专业学位）， 2022， 硕士

【摘要】 在深度学习领域中,卷积神经网络已经在图像分类、人脸识别、目标检测和语音识别等各种计算机视觉及自然语言处理任务中得到广泛应用。卷积神经网络的推理结果依赖于大量的网络参数及这些参数参与的复杂计算,这会限制它们在计算能力与存储空间有限的设备上的应用。因此,针对卷积神经网络的加速对人工智能的普及应用具有重要意义。静态网络在计算时计算图与网络参数都是固定的,这可能会限制它们的表示能力、效率以及可解释性。与静态网络方法不同,动态网络方法可以在推理时动态改变网络结构和参数量,从而适应不同的输入。因此,使用动态执行的方法来加速卷积神经网络的研究不断兴起。为了在推理过程中实现动态网络结构或参数调整,现有的动态网络加速方法大多需要对原有的网络结构进行较大的修改,或者引入额外的参数训练。针对以上问题,本文提出了基于卷积池化特征的动态网络推理加速方法,并通过实验对其有效性进行了相关的验证,本文的主要工作包括:(1)分析了卷积神经网络中卷积-池化结构特征,发现可以通过提前确定池化的结果减少卷积中的冗余计算,从而确定了基于卷积池化特征进行动态推理加速的可行性。(2)设计了卷积核通道重要性的评估标准,根据该评估标准对卷积核通道进行排序,改变卷积通道的执行顺序,从而跳过一些不必要的卷积计算,提升推理速度。(3)在通用处理器平台和专用加速器平台对以上方法进行了实验验证,实验结果表明该方法相比于传统卷积网络方法能够在保证计算精度的同时,最多减少40%的参数计算量。由于本论文提出的方法只是对传统的卷积计算进行了输入通道的排序,并不会在训练阶段改变网络结构,可以在不进行额外训练的同时实现动态推理加速,为卷积神经网络的加速提供了新的思路。更多还原

【Abstract】 In the field of deep learning,convolutional neural networks have been widely used in various computer vision and natural language processing tasks such as image classification,face recognition,object detection,and speech recognition.The inference results of convolutional neural networks rely on a large number of network parameters and the complex computations in which these parameters participate,which limits their application on devices with limited computing power and storage space.Therefore,the acceleration of convolutional neural networks is of great significance to the popularization and application of artificial intelligence.Static networks have fixed computational graphs and network parameters at computation time,which may limit their representational power,efficiency,and interpretability.Unlike static network methods,dynamic network methods can dynamically change the network structure and amount of parameters at the inference stage,thus adapting to different inputs.As a result,research using dynamically executed methods to accelerate convolutional neural networks keeps emerging.In order to achieve dynamic network structure or parameter adjustment in the inference stage,most of the existing dynamic network acceleration methods need to make major modifications to the original network structure or introduce additional parameter training.Aiming at the above problems,this thesis proposes a dynamic network inference acceleration method based on convolution and pooling features,and verifies its effectiveness through experiments.The main work of this thesis includes:(1)The convolution-pooling structural features in convolutional neural networks are analyzed,and it is found that redundant computation in convolution can be reduced by determining the pooling results in advance,thus confirming the feasibility of dynamic inference acceleration based on convolutional and pooling features.(2)The evaluation criteria for the importance of the channels corresponding to the convolution kernels are designed.According to the evaluation criteria,the channels corresponding to the convolution kernels are sorted and the execution order of the convolution channels is changed,thereby skipping some unnecessary convolution computations and improving the inference speed.(3)The above method is experimentally verified on a general-purpose processor platform and a dedicated accelerator platform.The experimental results show that the method can reduce the amount of parameter computation by up to 40% compared with the traditional convolutional network method while ensuring the inference accuracy.Since the method proposed in this thesis only sorts the input channels of the traditional convolutional computation,and does not change the network structure during the training stage,dynamic inference acceleration can be achieved without additional training,which provides a new idea for the acceleration of convolutional neural networks.更多还原