hustliujian

> Can you please upload the modified codes in your repo ? I did it in company, it's not convient to upload code.

vgg19's weight was first load in [__init__](https://github.com/makeuptransfer/SCGAN/blob/5f5018969c33a4e309c82f273b88ddb0ba3a17f0/models/SCGen.py#L154), then self.SCGen was initialized by [line 63](https://github.com/makeuptransfer/SCGAN/blob/5f5018969c33a4e309c82f273b88ddb0ba3a17f0/models/SCGAN.py#L63), so it will be overwritten.

@kriscpu I find the matter is [self.D_optimizer.step()](https://github.com/lcy0604/EraseNet/blob/673bfad49ec5a09abee7ff6900a2a9e6e061e058/loss/Loss.py#L61) modifies the weights, but D_fake is used in GLoss computation later. So delete [D_fake = -torch.mean(D_fake)](https://github.com/lcy0604/EraseNet/blob/673bfad49ec5a09abee7ff6900a2a9e6e061e058/loss/Loss.py#L57C1-L57C62), and compute it like `D_fake = -self.discriminator(output,...