Closed zengyh1900 closed 6 years ago
The encoder uses a progressively growing structure, so it first outputs 4x4, then 8x8, etc., all the way to 256x256. Please check Progressive GAN for more detail!
@jerryli27 Thank you for your quick response very much. But did you mean decoder? In PGGAN, it has just generator and discriminator with mirrored structure. I wonder in your framework (I'm sorry that I didn't find out the framework illustration in your paper), does the encoder take image of original size as input then map it into latent vector in a fixed structure, Or does the encoder take 4x4, then 8x8, etc., downsampled image as input and map it into latent vector?
I mean encoder, not decoder or discriminator.
Oh sorry I misunderstood. The encoder does grow along with decoder and discriminator. The encoder basically shares the same structure as the discriminator. I didn't add an illustration in the current version of the paper, but you can check either the appendix or the source code for the network structure.
@jerryli27 oh I see. Thank you for your kindly help very much :)
Thanks for your work ,it's really interesting. btw, I wonder is your encoder fixed and linear training, or in a progrresive growing? thx a lot!