search

StevenShaw1999 / RSSA

54 stars 2 forks source link

Training (adapting) my own GAN errors #3

Closed arith3 closed 2 years ago

arith3 commented 2 years ago

Hello.

First of all, thanks for releasing the code. It will be veryvery useful.

I found some issues when training my own images.

[invert_gan.py] When I want to use a target image more than 5, I have to manually modify range() in line 133. Can you fix this through the args function?

[train_????_proj.py] What does the 'assert args.source_key in args.ckpt' in line 775 mean? I delete out the line after make sure it wasn't used.

[model.py] 'latentcode.npy' in the 'exps' directory is used for the existing code, but I checked that the calculated result in 'invert_gan.py' is stored in latent//latent. So, in 'exp_name' of the init part of 'class Projection_module', I added the code in the same way as latent_dir = 'latent//latent'. Also, the part that does np.load has been changed appropriately. {np.load(f'%s/{i}_latentcode.npy'%latent_dir)}

I'm training my code like this, and I'm wondering if it's the same as your intended function. Also, could you please edit the code later?

Thanks. Have a nice day!

StevenShaw1999 commented 2 years ago

Hello! Thank you for your reminder! 1 I have modified the code in invert_gan.py. Sorry for that. 2 It's ok to delete this option, it is used in the few-shot-gan-adaption to ensure that you are adapting your model from the right source domains. 3 Thank you for this. I found that I haven't pushed the model.py file to the git after I modified it on the remote server. I have uploaded it just now. You can check it now.

arith3 commented 2 years ago

Thanks so much for the quick response and update!