I think Transfer learning with tensorflow 1.14 is something, there is not much clarity. In keras, while using pretrained model, we can freeze backbone weights in initial epo