赞
踩
初始的学习率为lr,可以在前面的0.8*num_epochs,学习率下降到 lr*0.1,然后在0.9*num_epochs下降到 lr*0.01;
也可以多次下降,0.4642,下降6次,0.4642**6=0.010005316163952237,0.4642**3=0.100026577288