Pick a safe batch size to avoid OOM in Tensorflow 2

Question

I have an outer loop that picks a pretrained imagenet model (MobileNet, ResNet, etc) and fine-tunes it. I would like the models to train as fast as possible which means I wa