Week4: Neural style transfer- layers choice in style cost function

Hi, I am aware that when we choose the layer from which we get activation to compute style cost has to be niehter too shallow nor too deep. also, I know that we get a better result if we merge the output of several layers.
my question is how do we choose these layers (based on what criteria)?
or it can be considered as a hyperparameter that we can try different values and choose the best ( of course the layers would be chosen niether too shallow nor too deep).

Thank you for your help.

It’s a good question. They don’t really give a hard and fast rule in the notebook or in what Prof Ng says in the lectures. My interpretation of what they say in the notebook is the way you phrased it:

It is an interesting experiment that you can try starting from choices they gave us in the notebook. Try some different combinations and see how that affects the results. Of course this particular case is just a matter of aesthetics, meaning that there is no “correct” answer: it’s just what looks good or interesting to you and that is to some degree a matter of personal taste.

Thank you for your response. Glad to know that my intuition was right. :nerd_face: