TensorFlow 2 Pocket Reference

Chapter 4. Reusable Model Elements

Developing an ML model can be a daunting task. Besides the data engineering aspect of the task, you also need to understand how to build the model. In the early days of ML, tree-based models (such as random forests) were king for applying straight-up classification or regression tasks to tabular datasets, and model architecture was determined by parameters related to model initialization. These parameters, known as hyperparameters, include the number of decision trees in a forest and the number of features considered by each tree when splitting a node. However, it is not straightforward to convert some types of data, such as images or text, into tabular form: images may have different dimensions, and texts vary in length. That’s why deep learning has become the de facto standard model architecture for image and text classification.

As deep-learning architecture gains popularity, a community has grown around it. Creators have built and tested different model structures for academic and Kaggle challenges. Many have made their models open source so that they are available for transfer learning—anyone can use them for their own purposes.

For example, ResNet is an image classification model trained on the ImageNet dataset, which is about 150GB in size and contains more than a million images. The labels in this data include plants, geological formations, natural objects, sports, persons, and animals. So how can you reuse the ResNet model to classify ...