Transfer Learning

In the previous chapter, you got hands-on practice training different CNN models for image classification purposes. Even though you achieved good results, the models took quite some time to learn the relevant parameters. If you kept training the models, you could have achieved even better results. Using graphical processing units (GPUs) can shorten the training time, but it will still take a bit of time, especially for bigger or more complex datasets.

Deep learning researchers have published their work for the benefit of the community. Everyone can benefit by taking existing model architectures and customizing them, rather than designing architectures from scratch. More than this though, researchers also share the weights of their models. You can then not only reuse an architecture but also leverage all the training performed on it. This is what transfer learning is about. By reusing pre-trained models, you don't have to start from scratch. These models are...