Pytorch Note57 Pytorch Visual network structure

A summary of all notes ： Pytorch Note Happy planet

With the development of deep neural network , The structure of the network is becoming more and more complex , It is also difficult to determine the input structure of each layer , Output structure, parameters and other information , This makes it difficult for us to finish in a short time debug. Therefore, it is necessary to master a tool that can be used to visualize the network structure . Similar functions are in another deep learning library Keras You can call one called model.summary() Of API It's easy to implement , After calling, our model parameters will be displayed , Input size , Output size , Overall parameters of the model, etc , But in PyTorch There is no such convenient tool to help us visualize our model structure .

about pytorch Come on , The visualization of model structure is still more important , This will facilitate our understanding of the data , And it can also deepen the understanding of the convolution change of each layer of data . Today's article briefly introduces , Visualization of some models , It's what I usually use to write code , It can also be used to check whether the code can output correctly .

Use print Print

In fact, the simplest thing is to use print Print , For example, we don't understand the official implementation of one of the networks , We can torchvision Import our model

I simply use torchvision Medium alexnet Model to test

from torchvision import models
net = models.alexnet()
print(net)

Then use it directly print Print , We can directly see the parameters of the internal implementation , We can use this to have a better understanding of our network model , We can also use these models for transfer learning , Just change the last level of classification .

AlexNet(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
(1): ReLU(inplace=True)
(2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
(3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
(4): ReLU(inplace=True)
(5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
(6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(7): ReLU(inplace=True)
(8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(9): ReLU(inplace=True)
(10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
(classifier): Sequential(
(0): Dropout(p=0.5, inplace=False)
(1): Linear(in_features=9216, out_features=4096, bias=True)
(2): ReLU(inplace=True)
(3): Dropout(p=0.5, inplace=False)
(4): Linear(in_features=4096, out_features=4096, bias=True)
(5): ReLU(inplace=True)
(6): Linear(in_features=4096, out_features=1000, bias=True)
)
)

But simple print(model), Only the information of basic components can be obtained , It can't show the of each layer shape, The size of the corresponding parameter quantity cannot be displayed ,

torchinfo visualization

In fact, a lot of what I used before were torchsummary, But it seems that ,torchsummary and torchsummaryX It hasn't been updated for a long time , and torchinfo By torchsummary and torchsummaryX The reconstructed Library .

And for ,torchsummary Sometimes it looks a little bloated , Output dimensions and quantities of all layers , The deep network structure is a little bloated .

install torchinfo perhaps torchsummary

This is very simple , Is the use pip Can be installed , Open the command line , Enter the following command to install

pip install torchinfo torchsummary

Use torchinfo

Whether it's for us torchinfo still torchsummary Come on , We all use the library summary function , However, the two parameters are somewhat different

You can also choose your preferences according to your preferences summary

First, we can use our torchinfo Of summary function

from torchvision import models
net = models.alexnet()
from torchinfo import summary
summary(model, (1, 3, 224, 224)) # 1：batch_size 3: The number of channels in the picture  224:  The height and width of the picture 

torchinfo Provides more detailed information , Include

• Module information （ The type of each floor 、 Output shape And parameter quantities ）
  

• The parameter quantity and size of the whole model
  

• Memory size required for a forward or reverse propagation, etc
  

We can also see the previous summary function , For this part , Namely Layer The visualization of is different , This part of visualization also gives many parameters , But for the results of complex models , It won't be clear

from torchsummary import summary
summary(net, (3, 224, 224)) # 3: The number of channels in the picture  224:  The height and width of the picture 

Be careful ：
When using colab perhaps jupyter notebook when , Want to implement this method ,summary() It must be the unit （ namely notebook Medium cell） The return value of , Otherwise we need to use print(summary(...)) To visualize .