Deep Learning with PyTorch

Classifying dogs and cats using transfer learning

Transfer learning is the ability to reuse a trained algorithm on a similar dataset without training it from scratch. We humans do not learn to recognize new images by analyzing thousands of similar images. We, as humans, just understand the different features that actually differentiate a particular animal, say a fox from a dog. We do not need to learn what a fox is from understanding what lines, eyes, and other smaller features are like. So we will learn how to use a pre-trained model to build state-of-the-art image classifiers with very little data.

The first few layers of a CNN architecture focus on smaller features, such as how a line or curve looks. The filters in the later layers of ...