I have implemented YOLO v1 and train/tested it on synthetic images with geometric forms. Implementing the loss function thought me a lot on how backpropagation really works. I used keras/tf.