Most efficient way to manually load large number of images from many subdirectories?

Question

I have a large image dataset with 3000 folders that I am attempting to create a training, test, and validation set for, which has the following structure:

ETL9G_I