Ultra-Light-Fast-Generic-Face-Detector-1MB icon indicating copy to clipboard operation
Ultra-Light-Fast-Generic-Face-Detector-1MB copied to clipboard

Published 20 hours ago • verified publisher icon Linzaer

The performance decreases significantly when convert the pytorch model to MNN model.

Open zhangxilun opened this issue 4 years ago • 1 comments

I use version-RFB-320.pth as pretrained pytorch model and validate the model on wider_face val set. And the performance is Easy AP:0.786, Medium AP: 0.696, Hard AP: 0.435. Then I convert the pytorch model to MNN model according to what you mentioned in Readme.md. However, I find the performance is Easy AP: 0.738, Medium AP: 0.596, Hard AP: 0.311 when using MNN. Why the performance decreases so dramatically just because I convert the pytorch model to the MNN model?

zhangxilun avatar Aug 04 '20 07:08 zhangxilun

I use version-RFB-320.pth as pretrained pytorch model and validate the model on wider_face val set. And the performance is Easy AP:0.786, Medium AP: 0.696, Hard AP: 0.435. Then I convert the pytorch model to MNN model according to what you mentioned in Readme.md. However, I find the performance is Easy AP: 0.738, Medium AP: 0.596, Hard AP: 0.311 when using MNN. Why the performance decreases so dramatically just because I convert the pytorch model to the MNN model?

My problem is the same as yours, have you solved it?

fengyanWang avatar Sep 28 '20 08:09 fengyanWang