In the “Cost Style” video of the “Neural Style Transfer” section of week 4 professor Ng says it seems redundant to use 2 different hyperparameters, alpha and beta, as coefficients for the content and style cost functions (equation: J(G) = alphaJ-content(C,G) + betaJ-style(S,G)). It actually makes sense to me that we would use two different hyperparameters here since the fact that they are distinct allows us to adjust alpha and beta separately depending on whether we want our generated image to have more fidelity to the style of image S or the content of image C. I’m curious if that intuition makes sense to anyone else.
I think your intuition makes sense. The two are independent and it makes sense to be able to tweak the relative weights. Maybe what Prof Ng meant is that if all you cared about was the relative influence of each cost, you could control that with just one parameter multiplying one or the other of them, but I think having two separate coefficients seems a bit more natural.