search

TArdelean / CainGAN

Other
53 stars 8 forks source link

About CainGAN w/o T (1) setting #2

Closed kenoharada closed 4 years ago

kenoharada commented 4 years ago

Thank you very much for sharing awesome work!

I would like to reproduce the result of CainGAN w/o T (1) (top row of table 2 in the paper) (I know the selected videos may change, so I can't exactly reproduce the result)

Could you tell me the hyperparameters for that and is there any additional setting for that?

TArdelean commented 4 years ago

The hyperparameters used are the ones in options.py (default), except for K=1 (since you want to reproduce 1 shot results). Note that CainGAN w/o T has a slightly different design since the input has fewer channels. You can't currently switch to that version with a parsable argument. You can, though, modify the code so that the target landmark is not used in the embedder: in_t=3 becomes in_t=0 and you don't use target_landmark in forward method.

kenoharada commented 4 years ago

Thank you very much for quick response! I modified the network architecture