RT. The fake_unet serves as a score function for generated data, and its weights are updated using Denoising Score Matching (DSM), implemented in "compute_loss_fake".
When the "--gan_alone" flag is set, the discriminator's loss consists solely of the guidance classification loss, which is implemented in "compute_guidance_cls_loss". However, even in GAN-alone mode, the final loss still includes both "loss_fake_mean" and "guidance_cls_loss", as they are summed together in the computation.
RT. The fake_unet serves as a score function for generated data, and its weights are updated using Denoising Score Matching (DSM), implemented in "compute_loss_fake".
When the "--gan_alone" flag is set, the discriminator's loss consists solely of the guidance classification loss, which is implemented in "compute_guidance_cls_loss". However, even in GAN-alone mode, the final loss still includes both "loss_fake_mean" and "guidance_cls_loss", as they are summed together in the computation.