iPERDance / iPERCore

Liquid Warping GAN with Attention: A Unified Framework for Human Image Synthesis
https://iperdance.github.io/work/impersonator-plus-plus.html
Apache License 2.0
2.42k stars 311 forks source link

adversarial generator loss #135

Open orydatadudes opened 2 years ago

orydatadudes commented 2 years ago

thank you for your code i have a question about the adversarial loss part

lwg_trainer.py -> optimize_G line 767 the loss is calculate by match the output to zero " self._loss_g_adv = self.crt_gan(d_fake_outs, 0) * self._train_opts.lambda_D_prob" for my understanding, d_fake_outs is 6 channels : the first 3 channels are generate target the last 3 channels are the ground truth pose(or shape) of the target person my question is why not match it to 1? instead of 0 thank you