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

compressed horse2zebra model #3

Open 0russ opened 4 years ago

0russ commented 4 years ago

Hello, thanks for sharing the code. I trained the cycleGAN model on horse2zebra dataset for several times without the quant. But the best FID is about 97, I wonder if I should change some hyperparameters when training the model without quant? And could you please upload the horse2zebra model as a reference， thank you very much!

htwang14 commented 3 years ago

Hi, thanks for your interest in our work! I have released the horse2zebra models compressed by GS32 and GS8. Please kindly check our GoogleDrive link.

On horse2zebra dataset, it is necessary to initialize the generator and discriminator with pretrained dense models. Please check the updated README section for more details. (On winter2summer_yosemite dataset, use dense model or random initialization get similar results so we just use random initialization for simplicity.)

I somehow forgot to include the dense initialization parts in our previous codes, and that's probably the reason you get higher FID scores. This issue is fixed with the new commits, and you should get similar performance as reported in the paper if you set rho between 0.008 and 0.009 and all other parameters at the default values using the latest codes.