[pytorch] fine tuning technology

Preface

Training neural networks is a very time-consuming thing , It requires a lot of computing power and a lot of data . Obviously, starting from scratch is not a wise choice , Making good use of existing resources is the wise choice .

Fine tuning technology

Generally speaking, image recognition can be divided into two steps ：

1. Extract the features of the picture , This part often passes CNN Convolution neural network .
2. According to the extracted features , To classify , This part is often realized by fully connected neural network .

• Is there any similarity between identifying a cat and identifying a dog ？
  The answer is yes , They are Extract picture features Are very similar .
  consider CNN The role of convolution in , It is identifying different edges , So whether it's a cat , Or dogs , Picture features are similar , But how to learn according to these characteristics is the key .

Pre training model

In image recognition , There are many classic Neural Networks , for example vgg,resnet etc. , For these classic Networks ,pytorch They all provide good training models . Some of these are in ImageNet On the training , With high accuracy . Use some trained to extract image features , It can greatly reduce the time of training .

Code implementation

import torch
from torch import nn
from torch.nn import functional as F
from torchsummary import summary

net = torchvision.models.resnet18(pretrained=True)
    
net.fc = nn.Linear(net.fc.in_features, 5)
nn.init.xavier_uniform_(net.fc.weight)
summary(net , input_size=(3,224,224) , device="cpu")

lr = 0.0005
loss = nn.CrossEntropyLoss(reduction="mean")

params_1x = [param for name, param in net.named_parameters()
    if name not in ["fc.weight", "fc.bias"]]
trainer = torch.optim.SGD([{
    'params': params_1x},{
    'params': net.fc.parameters(),'lr': lr * 80}],lr=lr, weight_decay=0.001)
epochs = 15

It's very simple , It's even simpler than defining neural networks completely by hand , Because it does not need to define its own network structure .
But these and training models can't be used directly , Some modifications are needed ：

1. Modify the last category number
   stay ImageNet in , The final full connection layer is an output of $1000$ Vector , That means $1000$ Categories , In practice, , It needs to be modified according to the number of categories recognized at present .
2. Learning rate
   Generally speaking , Pre trained parameters need not be modified , It can be set as a parameter without learning , You can also set its learning rate to be very small . And for the final full connection layer , That is, the network that classifies the extracted image information , Its learning rate will be relatively large .

Function and significance

Fine tuning technology is a blessing for people who don't have enough computing power , It greatly reduces the cost of training .

1. Extracting image features CNN Networks tend to be closer to input , Training through gradient back propagation , It often takes more time than a fully connected network close to the output .
2. Using micro carving technology , You can get a very good result in a very short time , stay 5 After a round of iterations, we can already achieve 91% The accuracy of ;15 One round of iteration can achieve 97% The accuracy of .
   

Shortcomings and limitations

Not all cases can use fine-tuning Technology , When there are significant differences in picture features , Using fine-tuning technology often does not get satisfactory results .
such as , stay ImageNet The models trained on are all based on normal pictures , But if it is used to recognize medical images （X Light film, etc ） It will lead to failure . Similarly, using this model to recognize cartoon pictures will often recognize .