xxh96 / ADCSR

DSADCSR FOR AIM2019 Extreme Super-Resolution Challenge - Track 1: Fidelity
13 stars 1 forks source link

为什么代码里在卷积层前用了wn,论文中4.4初始化lr是1e-4不是1e-3? #9

Closed splinter21 closed 4 years ago

splinter21 commented 4 years ago

如果按照wdsr论文里的,学习率应该是10倍?

xxh96 commented 4 years ago

wn参考WDSR,lr你可以自己决定,在将输入归一化到-1~1之间时,通常都是在1e-4这个数量级开始训练,参考EDSR,RCAN等.