Transfer learning - getting started

We usually use in-depth learning, such as VGG16 There will be a parameter in ,pretrain=True, This means that the model adopts the pre training model , The most commonly used method of transfer learning in pre training

A typical example is in Imagenet Train one backbone, Then on another new data set （ such as cifar、cub） Training set tuning （fine-tune）backbone, Then test the model on the test set of this new data set .

Then why don't we start from scratch on the new data train A model ？ We all know ,Imagenet There are a lot of pictures , And the picture covers a more comprehensive area , It can be approximately regarded as a description of the distribution of real-world data , So hope in ImageNet The model trained on can extract general picture features , This common feature is likely to migrate to a picture domain not seen downstream . Therefore, it is widely believed that , stay ImageNet（ Or a larger data set ） Train one backbone, Then fine tuning is the best way .

transfer learning There is one difference from domain adaptation The key point of , That is, the images of the data set during training and the data set during fine-tuning are not only domain Different ,category It's usually different . because category Different , As a result, the original network classification layer cannot be used during fine-tuning , I have to learn another ; And because the domain It's different. ,backbone The extracted features are not enough discriminative 了 , Therefore need finetune backbone. You will see , From these two points, it will go straight to few-shot learning Core issues .

（ About domain and category Please see the introduction of small sample learning ）
 

Transfer learning classifies according to sample labeling