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

为何mse会不如感知loss? #2

Closed lishiyuwhu closed 3 years ago

lishiyuwhu commented 4 years ago

看了论文和gs.py, 给我的感觉是是基于cyclegan生成的图像, 外加剪枝loss重新训练一个pix2pix的G. 我这样的理解有问题吗?

既然是衡量student_output_img和teacher_output_img的差距, 为什么感知loss是强于mse的呢? 我的理解是训练目标就单纯是让G_p2p(student)的结果拟合G_cyclegan(teacher)的结果, mse这样直接的约束为何没有感知loss这样更加偏向语义的约束效果好呢?

htwang14 commented 3 years ago

Hi, thanks for your interest in our work! MSE loss mainly focuses on pixel-wise differences, ignoring global content and style features which are more desired in style transfer tasks. In contrast, perceptual is better at capturing global content and style features. Please refer to the original perceptual loss paper for more details.