Microsoft computer vision best practice guidelines using fastai/pytorch lib

I have just found an interesting repo from Microsoft which covers best practices in computer vision, from training model to deployment.
They are importing fastai library for 2 tasks - image classification and image similarity - while object detection uses PyTorch vision library.
It might be a good place to see a full pipeline from training to deploy a fastai/PyTorch model to production. They also describe hard-negative sampling technique which I find very useful in real applications.