what is LSNCD in the code? is it wnorm?

[john09282922]

Hi, When I see your code, I can't find the LSNCD and How did you make style space in your code?

[caopulan]

loss:

GANInverter/training/encoder_trainer.py

Line 332 in 4333561

if self.opts.sncd_lambda > 0: # calculate cos loss though lambda=0

calculate $\mu$:

GANInverter/training/encoder_trainer.py

Line 80 in 4333561

if self.opts.sncd_lambda > 0:

[caopulan]

Some codes are not organized well now, which may be confusing to read. But scripts (training/inference) and configs of each method are available to reproduce the experiments. Have fun with GANInverter!
We are going to clean and re-organize the codes (v2.0) and release the weight after our paper is accepted.

[john09282922]

thank you for replying this comment quickly, are you still working on SAM, right? In here, is there no implementation code? I am wondering, In SAM github, they did not have a code for encoder, only optimization, right?

[caopulan]

We are working on the implementation of SAM/HFGI/Hyperstyle now, especially their inference phase. SAM may be implemented in GANInverter in February or March.
Yes, they only release the optimization codes. I report an issue to ask whether they will release the encode but have not received any replies.
However, if they don't release their official encoder codes, we may not implement them in GANInverter, because we are working on a new inversion method, "Domain-Specific Hybrid Refinement", which gains significant improvement to all previous SOTAs. But it is only an optimization-based method. We may study its encoder version later.

[john09282922]

thank you so much. I have another question, I want to know how to get style space from w space. and do you try to use w+space? do you know how to get w+space from w?