[paper notes] effective CNN architecture design guided by visualization

The paper

 

Preface

Modern efficient Convolutional Neural Networks (CNN) Always used Depth separates the convolution  (DSC) and Neural network architecture search  (NAS) To reduce the number of parameters and computational complexity . But it ignores some inherent characteristics of the network . Subject to visual feature map and N×N(N>1) The inspiration of convolution kernel , This article introduces several guidelines , To further improve parameter efficiency and reasoning speed .

The parameters designed based on these guidelines are efficient CNN The architecture is called  VGNetG, Better accuracy and lower latency than previous Networks , The parameters are reduced by about 30%~50%.VGNetG-1.0MP  stay  ImageNet  Classify the data set with  0.99M  The parameters of are implemented  67.7%  Of  top-1  Accuracy rate , stay  1.14M  Under the parameters of  69.2%  Of  top-1  Accuracy rate .

Besides , prove Edge detector It can be fixed by edge detection kernel Replace  N×N kernel To replace the learnable deep convolution layer to mix features .VGNetF-1.5MP  achieve  64.4%(-3.2%) Of  top-1  Accuracy and accuracy  66.2%(-1.4%) Of  top-1  Accuracy rate , With additional gaussian kernel.

 

1 Methods of this paper

The author mainly studies the..., which is constructed by standard convolution 3 A typical network ：

1. Standard convolution ==>ResNet-RS,

2. Group convolution ==>RegNet,

3. Depth separates the convolution ==>MobileNet、ShuffleNetV2 and EfficientNets.

These visualization results show that ,M×N×N kernel There are obviously different patterns and distributions in different stages of the network .

1.1 CNN You can learn how to satisfy the sampling theorem

Previous work has always believed that convolutional neural networks ignore the classical sampling theorem , But the author found that the convolutional neural network can meet the sampling theorem to a certain extent by learning the low-pass filter , Especially based on  DSCs  Network of , for example  MobileNetV1  and  EfficientNets, Pictured 2 Shown .

1、 Standard convolution / Group convolution

Pictured 2a and 2b Shown , Throughout M×N×N individual kernel There are one or more significant N×N individual kernel, For example, blur kernel, This phenomenon also means that the parameters of these layers are redundant . Please note that , remarkable kernel It doesn't necessarily look like Gauss kernel.

2、 Depth separates the convolution

Strided-DSC  Of kernel Usually similar to Gauss kernel, Including but not limited to  MobileNetV1、MobileNetV2、MobileNetV3、ShuffleNetV2、ReXNet、EfficientNets. Besides ,Strided-DSC kernel The distribution of is not Gaussian , It's a Gaussian mixture .

 3、 The last convolution layer Kernels

modern CNN Always use the global pooling layer before the classifier to reduce the dimension . therefore , A similar phenomenon also appears on the final depth convolution layer , Pictured 4 Shown .

These visualizations indicate that depth convolution should be selected in the lower sampling layer and the last layer instead of standard convolution and group convolution . Besides , Fixed Gauss can be used in the lower sampling layer kernel.

1.2 Reuse feature maps between adjacent layers

Identity Kernel And similar characteristic graphs

As shown in the figure above , Many deep convolution kernels have large values only in the center , Just like the identity core in the middle of the network . Because the input is only passed to the next layer , Therefore, convolution with identity kernel will lead to repetition of characteristic graph and computational redundancy . On the other hand , The following figure shows that many feature maps are similar between adjacent layers （ Repetitive ）.

therefore , Partial convolution can be replaced by identity mapping . otherwise , Depth convolution is slow in early layers , Because they are usually not fully utilized  Shufflenet V2  Modern accelerators reported in . So this method can improve parameter efficiency and reasoning time .

1.3 Edge detector as a learnable depth convolution

Edge features contain important information about the image . As shown in the figure below , Most of the kernel Similar to edge detection kernel, for example  Sobel filter kernel  and Laplace filter kernel. And this kernel The proportion of is reduced in the later layer , And like ambiguity kernel Of kernel The proportion increases .

therefore , Maybe Edge detector Can replace based on  DSC  Deep convolution in the network of , To mix features between different spatial locations . The author will use edge detection kernel Replacement can be learned kernel To prove that .

 

2 Network architecture

2.1 DownsamplingBlock

DownsamplingBlock  Halve the resolution and expand the number of channels . Pictured a Shown , Only extended channels are generated by pointwise convolution to reuse features . The kernel of deep convolution can be initialized randomly or use a fixed Gaussian kernel .

2.2 HalfIdentityBlock

Pictured b Shown , Replace half depth convolution with identity mapping , And reduce while maintaining the block width half pointwise convolutions.

Please note that , The right half of the input channel becomes the left half of the output channel , In order to better reuse features .

2.3 VGNet Architecture

Use  DownsamplingBlock  and  HalfIdentityBlock  Constructed a  VGNets. whole  VGNetG-1.0MP  The architecture is shown in the table 1 Shown .

2.4 Variants of VGNet

For further study N×N The impact of the kernel , Introduced  VGNets  Several variants of ：VGNetC、VGNetG  and  VGNetF.

VGNetC： All parameters are randomly initialized and learnable .

VGNetG： except  DownsamplingBlock  Outside the kernel of , All parameters are randomly initialized and learnable .

VGNetF： All parameters of depth convolution are fixed .

 

3 experiment

Reprint ： Lightweight Backbone overlord | VGNetG achievement “ No choice , All of them ” Lightweight backbone ！