5 Customizing the search space by creating AutoML pipelines

In chapter 4, we solved a variety of problems with AutoKeras without customizing the search space. In practice, you may want to use a specific ML algorithm or data preprocessing method to solve a problem, such as an MLP for a regression task. Designing and tuning a particular ML component requires tailoring the search space, tuning only the relevant hyperparameters while fixing some others.

This chapter introduces how to customize the search space based on your requirements and automatically discover certain kinds of deep learning solutions for different types of tasks. Constraining the search space can also reduce the search time, allowing you to achieve better results with fewer trials. You will learn how to customize the search space by creating both sequential and graph-structured AutoML pipelines. I’ll show you how to implement an AutoML pipeline with the AutoKeras functional API, and how to use the built-in blocks in AutoKeras to conduct automated hyperparameter tuning and model selection for different tasks. You’ll also learn how to customize your own building blocks when the pre-existing blocks does not satisfy your need.

5.1 Working with sequential AutoML pipelines

5.2 Creating a sequential AutoML pipeline for automated hyperparameter tuning

5.2.1 Tuning MLPs for structured data regression

5.2.2 Tuning CNNs for image classification

5.3 Automated pipeline search with hyperblocks

5.3.1 Automated model selection for image classification

5.3.2 Automated selection of image preprocessing methods

5.4 Designing a graph-structured AutoML pipeline

5.5 Designing custom AutoML blocks

5.5.1 Tuning MLPs with a custom MLP block

5.5.2 Designing a hyperblock for model selection

5.6 Summary