search

VITA-Group / GAN-Slimming

[ECCV 2020] "All-in-One GAN Compression by Unified Optimization" by Haotao Wang, Shupeng Gui, Haichuan Yang, Ji Liu, and Zhangyang Wang
MIT License
110 stars 23 forks source link

Confusion about dense model？ #6

Open XGuider opened 3 years ago

XGuider commented 3 years ago

Thank you best work,i meet some confusion about dense model.why need dense model? The dataset is already one-to-one. I think we can skip directly to the fourth step without a dense model, but I don’t know if my idea is correct. Hope you can help me.

XGuider commented 3 years ago

if i use gs.py for other network,whether i need change model?

htwang14 commented 3 years ago

Thanks for your interest in our work. GAN training is unstable due to the minimax optimization process, and combining it with model compression makes it even more unstable due to the use of non-differentiable sparse constraints. Using the dense model for distillation stabilizes the training process and empirically achieves better results. You can definitely directly compress GAN without the dense model, but it won't achieve as good a performance, as shown in our ablation study.

htwang14 commented 3 years ago

if i use gs.py for other network,whether i need change model?

Yes, you will need to replace CycleGAN (the model currently used) with your own network.