Learning Deep Learning: Theory and Practice of Neural Networks, Computer Vision, NLP, and Transformers using TensorFlow

Chapter 8

Deeper CNNs and Pretrained Models

In this chapter, we describe three convolutional neural networks (CNNs): VGGNet, GoogLeNet, and ResNet. Both VGGNet (16 layers) and GoogLeNet (22 layers) are from 2014 and were close to human-level performance on the ImageNet dataset. VGGNet has a very regular structure, whereas GoogLeNet looks more complex but has fewer parameters and achieved higher accuracy. In 2015, both of these networks were beaten by ResNet-152 consisting of 152(!) layers. However, in practice, most people have settled on using ResNet-50, which consists of “only” 50 layers. As a programming example, we show how to use a pretrained implementation of ResNet and how you can use it to classify your own images. The chapter ends with ...