Add a neural style transfer model with AdaIN

[chunxiangzheng]

[chunxiangzheng]

@LukeWood

[LukeWood]

Notes from Luke/Jake 1:1:

• get_adain_encoder(), get_adain_decoder() returns a tuple of encoder and decoder
• Internally, we can use AdaIn as a custom layer
• AdaIn(include_rescaling=..., weights=”mine.h5”) -> this is a functional model that just calls encoder()
• AdaIn() layer decoder()
• Internal that creates an AdaIn identical model from an encoder and decoder
• model = produceAdaIn(encoder, decoder) -> functional model
• Then we can call model.save(“mymodel.h5”)
• examples/models/training/style_transfer/adain.py -> define the model trainer the same way we have defined it before

[LukeWood]

As far as the API is concerned; lets wrap the inference logic in a top level class that loads the proper weights to the models, and implements inference in a style_transfer() method. You can see an example of this here: https://github.com/keras-team/keras-cv/blob/master/keras_cv/models/generative/stable_diffusion/stable_diffusion.py#L41

[tanzhenyu]

Thanks for the PR! I'm not sure about the context here. It's not aligned with our immediate roadmap, so if this is important, please go ahead and address the comments so we can merge it. Or we can close it for now, and if this becomes needed we can re-open it later.

[ianstenbit]

Hey @chunxiangzheng -- where did we land on this PR? If you're interested in finishing it up, please feel free to address comments and re-request a review from Luke.

I'm going to close this PR for now, but just re-open it if you'd like to keep it going!