8 Model Training and Validation: Part 1
This chapter covers
- Developing the Model Training and Validation components
- Capturing metrics and artifacts in tracking frameworks
- Adding the Model Training and Validation components to pipelines
- Different methods to access training and evaluation data
In the previous chapter, we've laid the groundwork to download the identity card dataset that includes the images and labels, process them into a format that YOLO expects, then divided them up into train, test and validation splits. Through that process, you've also learnt how to create your first Kubeflow components and pipelines.
The next step, and arguably the most fun, is model training and evaluation. In this chapter, we'll carry on extending the data preparation pipeline from the previous chapter and bolt on training and evaluation components. In your initial project, you'll develop an ID card object detection system utilizing the popular YOLO (You Only Look Once) algorithm. Having mastered this concept, you'll then apply similar techniques to design a movie recommendation system.
8.1 Training an Object Detection Model
When we left off at the previous chapter, we had six lists, containing the cross product of either train, test, validation splits of either file names of images or file names of lists. These lists should then be passed to the training component that we will outline in this chapter. The training component then: