search

mit-han-lab / gan-compression

[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Other
1.1k stars 148 forks source link

Question about distillation #4

Closed 07hyx06 closed 4 years ago

07hyx06 commented 4 years ago

take CycleGAN compression as an example.

the teacher generator: a MobileNet-based CycleGAN trained from scratch the student generator: a MobileNet-based distillation CycleGAN model the final generator: a fine-tuned sub network of student generator

Is my understanding correct?

Which means the original CycleGAN model(with normal conv) is not needed in compression algorthm?

lmxyy commented 4 years ago

Yes, you are correct. The CycleGAN model (with normal conv) is not needed in our compressing pipeline. But we have compared our final model with the original model (with normal conv) in our paper.

07hyx06 commented 4 years ago

Thx for your quick reply! one more question: What about the behavior of the teacher generator and the student generator before NAS? I remember MobileNet can decrease 9xFLOPs on classfication. What about the MobileNet-based generator CycleGAN computation compare to the original Resnet-based one?

lmxyy commented 4 years ago

With the same ngf, the MobileNet-style generator is 5.8x smaller takes 3.1x less computation than the original resnet-based generator.

lmxyy commented 4 years ago

Do you have any other questions? If not, I will close the issue.

07hyx06 commented 4 years ago

no more. just close :D