Groupby, Apply, and Aggregate

Summary

Both groupby‐aggregate and groupby‐apply patterns are useful in either exploratory data analysis to understand data properties or data preprocessing for machine learning feature engineering. Groupwise operation on dense tensor inputs is seen in many computer vision models. This chapter aims to implement Visual Transformer from scratch to illustrate how this idea is being used in an application. The transformer model is a powerful sequence representation model that is responsible for the subsequent advancements in natural language processing, computer vision, audio processing, and time series forecasting. Bucketization is a powerful feature engineering technique that helps data scientists and machine learning engineers build more efficient models. Both Tensorflow and PyTorch have dedicated operations and quantization integer representations to help the parameter quantization and model compression processes. The chapter provides an illustration of an example of segment‐wise summation operations.