Call the classic architecture and build the model based on the classic

 1 Call the classic architecture    

        Most of the time , We won't go from 0 To create our own architecture , Instead, choose a suitable architecture or idea from the classic architecture , On the classic architecture, the architecture is customized and modified according to the needs of data （ Yes, of course , We can only call what we have learned 、 And master the structure of the principle , Otherwise, we will have no way to start when revising ）. stay PyTorch in , Basically all classical architectures have been implemented , So we can go straight from PyTorch in “ Transfer Library ” To use . Unfortunately , Most of the libraries directly transferred out cannot meet our own needs , But we can still call PyTorch As the basis of our own architecture .

       PyTorch There is a very important and easy-to-use package in the framework ：torchvision, The package is mainly composed of 3 The sub package consists of , Namely ：torchvision.datasets、torchvision.models、torchvision.transforms. from torchvision Invoking the complete model architecture , These architectures are located in “CV Common models ” modular torchvision.models in . stay torchvision.models in . stay torchvision.models in , framework / The model is divided into 4 Large type ： classification 、 Semantic segmentation 、 object detection / Instance segmentation and key point detection 、 Video classification . For each type of Architecture ,models Both contain at least one parent class that implements the architecture itself （ Present as “ Hump type ‘ name ） And a subclass containing pre training functions （ All lowercase ）. For having different depths 、 In terms of architectures with different structures , It may also contain multiple subclasses .（ Parents and children ： The relationship between the basic skeleton and the basic implementation ）

        For only one architecture 、 There are no different depths of AlexNet Come on , The structure of the two classes is the same

m.AlexNet()     # Check what parameters need to be filled in ？

m.alexnet()      # take AlexNet The functions of the parent class are included in , Parameter input to the original schema is not allowed , But you can do pre training

        For residual Networks , A parent class is a class that implements a specific architecture , Subcategory is a class that has filled in the required parameters

m.ResNet()          # You can implement structures of different depths from this class

m.resnet152()      # The specific depth and parameters are locked , You can perform pre training on this class

        When actually using the model , We almost always call lowercase classes directly , Unless we want to modify the internal architecture on a large scale . As shown below , The way to call a class is very simple ：

 

        Due to different network architectures , Therefore, different architecture classes do not share properties or methods , We need to check the architecture ourselves before calling . We can modify the super parameter settings of some layers according to our own needs . 

      If we want to modify the classic architecture , We have to revise... Layer by layer . One layer of the convolutional network may be used for all subsequent layers An impact , Therefore, we often only input to the network 、 Fine tune the output layer , The middle tier of the schema will not be modified . However , Most of it Time sharing and full application of classical architecture can not meet our modeling needs , So we need to build our own architecture based on the classic architecture . 

2 Self built architecture based on classic architecture

        Almost all modern classic architectures are based on ImageNet Data sets 224x224 The size of the 、1000 Constructed by classification , Therefore, almost all classical architectures will have 5 Next sampling （ The pool layer or step size is 2 The convolution of layer ）. When our data set is large enough , We will give priority to using the classic architecture to run on the data set , But when our image size is small , We don't have to expand the image to 224x224 Size to fit the classic architecture （ It can be used transform.Resize Do it , However, the prediction effect after amplification and conversion may not be very good ）. This will not only increase the computational power requirements 、 The calculation time becomes longer , It may not be able to achieve good results . If possible , On smaller pictures , We hope to keep the status quo as far as possible to control the overall amount of calculation .

        For convolution Architecture , Change the of the characteristic graph Number of inputs and outputs The behavior of is only related to oneortwo layers , To change the Dimension of characteristic drawing The behavior of will affect the whole architecture . therefore , We usually learn from the classic architecture ” extract “ Part to use , There is also a small possibility from 0 Build your own new architecture . Suppose we are using something similar to Fashion-MNIST Size ,28x28 Data set of , On such a dataset , The only chance we may perform the next sampling is 2 Time , Once from 28x28 Dimensionality reduction 14x14, Another time is from 14x14 Dimensionality reduction 7x7. Such a data set is not very suitable for tens of millions 、 Classic architectures with hundreds of millions of parameters . under these circumstances , torchvision.models The built-in architecture cannot flexibly meet the requirements , Therefore, we often do not directly use the built-in Architecture , Instead, the architecture is reconstructed on the basis of its own architecture .

Here's an example based on VGG And residual network self built architecture ：

        We may have many different ideas for building architecture . The most common way is to follow VGG To deepen the network , The other is to use blocks in classical Networks （ For example, the classical residual unit in residual network 、GoogLeNet Medium inception Isostructure ） To deepen the network . Experience proves that , stay inception It would be good to add some ordinary convolution layer before the residual unit , So here I use VGG The idea of , Then use the residual unit . You can freely combine your favorite structures .