5 Transfer Learning: Reusing Pretrained Neural Networks

 

This chapter covers:

  • What transfer learning is and why it is better than training models from scratch for many types of problems
  • How to leverage the feature-extraction power of state-of-the-art pretrained convolutional neural networks (convnets) by converting them from Keras and importing them into TensorFlow.js
  • What SymbolicTensors are and how they help you achieve flexible “plug and play” of model components
  • Why you should let only some layers of the model update by freezing other layers during transfer learning
  • How to replace the output layer of a pretrained convnet with new output layers to solve different types of transfer-learning tasks and dataset
  • What is the fine-tuning technique is and how it helps you get more accurate models from transfer learning
  • How to use transfer learning to achieve object detection in TensorFlow.js

5.1  Introduction to transfer learning: Reusing pretrained models

5.1.1  Transfer learning based on compatible output shapes: Freezing layers

5.1.2  Transfer learning on incompatible output shapes: Creating a new model using outputs from the base model 

5.1.3  Getting the most out of transfer-learning through fine-tuning: An audio example

5.2  Object detection through transfer learning on a convnet

5.2.1  A simple object detection problem based on synthesized scenes

5.2.2  Deep dive into simple object detection

5.3 Summary

5.4 Exercises