Tensorflow 2. Chapter 15 of X (keras) source code explanation: migration learning and fine tuning

1. Transfer learning and fine tuning

• The migration study Including acquiring features learned from a problem , Then apply these features to new similar problems . for example , Features from models that have learned to recognize raccoons may be useful for building models designed to recognize civet cats . For the task that there is too little data in the data set to train the complete model from scratch , Transfer learning is usually performed .
• In the context of deep learning , The most common forms of transfer learning are the following ：
  • from Before Get from the training model layer .
  • frozen These layers , To avoid destroying any information they contain in subsequent training rounds .
  • In already Add some new trainable layers to the top of the frozen layer . These layers learn to transform old features into predictions for new data sets .
  • On the dataset fine-tuning , Including thawing the whole model obtained above （ Or part of the model ）, Then retrain the model with a very low learning rate on the new data . Adapt the pre training features to the new data in an incremental way