Implement WnB for training visualization

[hermancollin]

I checked and implementing weights n biases is pretty straight forward for YOLO. Ideally we should log metrics and predictions to track the training. This should also be easy to add to retinanet