You are tasked with developing an efficient input pipeline for an ML training model that needs to process images from various sources quickly. The images are too large to fit into memory at once. Considering Google's recommended best practices, how should you go about creating this dataset?