Why use fixstep lr

[liu6381810]

Thanks for your work! I have 2 questions: In the paper, the author use cos lr but why fixsLR used in your code? And now W3A2 acc has reported but I want to know whether you train W4A4 resnet18 reaching the acc in the paper?

[zhutmost]

Ahhh, I cannot obtain better acc with cos decay, and I used step decay but fixed decay. And the config.yaml is just a template. Please refer the configuration file in the example folder.

I have tried W4A4, but its accuracy is less than the authors' one.

[liu6381810]

Thanks for your reply, Another quesiton I want to know is have you ever verified the effect of grad_scale In my early quantization experiment(not lsq) weight lr is often much smaller than scale lr but with grad scale it's approximate decay the lr for scale, it's confused for me

[zhutmost]

Sorry not yet.