Also note that 8K input images is a radically larger dataset than the ones they give us here in the courses. E.g. in the famous “cat/not cat” datasets back in Course 1, we had 209 training images and 50 test images. Training a neural network can be very time consuming precisely because it almost always takes a large training dataset to get good results. In fact you could legitimately wonder how we got such good results on the “cat recognition” problem with such a tiny dataset. Here’s a thread with some experiments to show that the 209 training + 50 test images are very carefully “curated” to get surprisingly good results with such a small training dataset.