arcface-pytorch icon indicating copy to clipboard operation
arcface-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon bubbliiiing

eval_LFW.py

Open chenqian57 opened this issue 1 year ago • 3 comments

为什么我下载了预训练好的权重arcface_mobilefacenet.pth以后，设置backbone为mobilefacenet，运行eval_LFW.py结果accuracy只有百分之60多呢

chenqian57 avatar Apr 17 '23 08:04 chenqian57

0 0不太合理，你在哪里下的0 0

bubbliiiing avatar Apr 17 '23 15:04 bubbliiiing

0 0不太合理，你在哪里下的0 0

不好意思刚看到，我是在你发的网盘https://pan.baidu.com/share/init?surl=ElJlfmMwOGX699MsgLY8qA提取码: z3rq里下载的权重（我没有执行训练，直接评估的）， 现在是把pytorch从1.10.2+cu102换成了1.12.0，arcface_mobilefacenet.pth和arcface_iresnet50.pth评估的accuracy是对的，但是mobilenetv1的评估结果不太对，显示Accuracy: 0.50000+-0.00000，Best_thresholds: 0.00000，Validation rate: 0.98767+-0.00517 @ FAR=0.97333，想问一下mobilenetv1除了backbone 和model_path要修改，还有别的地方要改吗

chenqian57 avatar Apr 18 '23 03:04 chenqian57

Loading weights into state dict... Test Epoch: [5888/6000 (96%)]: : 24it [00:12, 1.87it/s] Accuracy: 0.96200+-0.01090 Best_thresholds: 1.23000 Validation rate: 0.74900+-0.03792 @ FAR=0.00100

Process finished with exit code 0 偏低，这个是啥问题，下载你的权重

pengyang1225 avatar Oct 20 '23 07:10 pengyang1225