Part 3 is your guide to the know-how and practical insights needed to apply deep learning to tabular data problems. As a stand-alone solution or integrated with gradient boosting, deep learning can get good results with tabular data when you know how to use its unique way of finding solutions to predictive tasks.
Chapter 8 explores various deep learning stacks and frameworks for working with tabular data, including low-level frameworks like TensorFlow and PyTorch and high-level APIs like fastai and Lightning Flash. It introduces several libraries specifically designed for tabular deep learning tasks, such as TabNet, PyTorch Tabular, SAINT, and DeepTables. We compare the different stacks and discuss each one’s strengths and weaknesses. Chapter 9 extends the discussion to best practices. We use the Kuala Lumpur real estate dataset to illustrate these best practices for deep learning with tabular data, including data preparation, model architecture design, and model training. A Keras-based project, the example emphasizes easy, understandable, and effective data pipelines, along with a modular approach that promotes code reuse.