Junlin Han
Junlin Han
The default NCE loss sample patches from 5 layers while the first layer is actually RGB pixels.
> > > > The default NCE loss sample patches from 5 layers while the first layer is actually RGB pixels. > > Thank you very much for the reply....
No, it should be unpaired out. Train it again or adjust the weights of losses. ( I guess make lambda_GAN to 2 helps)
> > > @JunlinHan Thanks for the advice. I have looked at all of the training options. I don't understand what to change. Do you get an unpaired output? Yes,...
Chcek the mlp_sample in models/network.py . Patch are encoded to features. NCE loss is based on features.
Default NCE layer setting is based on Resnet-9 generator. You need to redesign and change layers.
OK, let me end this question. Hinge loss works unwell in some tasks like summer2winter, photo to VanGogh/Monet painting. Those tasks usually do not have a clear distinction between background...
I think it's for aligning the dataset Cat Dog( AFHQ(Animal faces dataset)). If you are interested in this dataset, you may go to StarGANV2 repo and download it. It's already...
You can modify the code to enable it. But the workload is quite big. Maybe a simple solution is to downsample the generated output to 64x64.
> > > @JunlinHan Hi Junlin, thanks for your reply. > You mentioned that modifying the code to fit different sizes of the input and output needs big workload. >...