ezioby/glead

[CVPR 2023] GLeaD: Improving GANs with A Generator-Leading Task

glead's Issues

Uploading code?

Hi,
Are you intending to upload your code ?

Default settings for training (combination with GGDR)

Hello,

Default settings in the "train_scripts_.py" uses reg_target_fake as 'gfeat' and reg_target_real as 'gfwimage'.
In this setting means the combination of glead and ggdr? ( in Table 1.)

Thanks,

Glead loss computed more than necessary in Dreg phase

In loss_glead.py in line 207. if do_Dmain or do_Dr1:. My understanding is that this if block is supposed to compute the GAN loss for the real images + the glead regularization loss and or the R1 regularization loss. And I guess that if only R1 regularization is wanted (meaning do_Dmain is False then neither the gan loss nor the glead regularization loss is supposed to be computed. While in fact the GAN loss is not computed (line 231 if do_Dmain: ), the glead loss is still computed one extra time (one more time than for fake images). So my question is whether this was or wasn't intentional.