search

jdnie / AdderNetCuda

CUDA实现'huawei-noah/AdderNet'的forward和backward
BSD 3-Clause "New" or "Revised" License
17 stars 0 forks source link

About the accuracy of the training #2

Open Guopeng1019 opened 4 years ago

Guopeng1019 commented 4 years ago

used your training code, the accuracy is about 50%，but the paper's accuracy has achieved 90%,could you please tell me why?

jdnie commented 4 years ago

I don't know why, I checked my codes and didn't find the bugs, maybe you can help me. But the addernet's grad transform is very difficult, the |W+X|'s grad is 1 or -1, it's too small and difficult to transform to the deeper layers.

StephenZhao1 commented 3 years ago

used your training code, the accuracy is about 50%，but the paper's accuracy has achieved 90%,could you please tell me why?

I have met the similar question.On my dataset,the accuracy is 100% at the very beginning.