In lesson 2, Jeremy says that when we have images which are not like the images on imageNet, we might need to make the learning rate 3x lower or similar.
My question is, for images like water bodies (sea, ocean, river) or like roads (highway, city roads, multi-lane super-ways )
These images are very similar in color, shape and structure and there is no central object to focus on. So what techniques can we use to train our model, should we still go for lower learning rates of unfrozen layers? or do we need to train a new model for this.