search

deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
23.33k stars 5.41k forks source link

The verification accuracy of arface_torch is not improving #1737

Open hhhThomas opened 3 years ago

hhhThomas commented 3 years ago

Hello,

I am training a mobilefacenet(embedding=128) with CASIA_Webface using arcface_torch. The loss has decreased from 53.7805 to 2.8381. However, the verification accuracies of lfw, cfp_fp, and agedb_30 remain 0.5, which does not make any sense. Could you please point out where I might go wrong? Thank you.

Training: 2021-09-03 04:05:58,634-Speed 322.19 samples/sec Loss 2.6971 LearningRate 0.0500 Epoch: 1 Global Step: 2000 Required: 14 h ours testing verification.. (12000, 128) infer time 44.320912999999905 Training: 2021-09-03 04:06:46,999-[lfw][2000]XNorm: 9.465224 Training: 2021-09-03 04:06:46,999-[lfw][2000]Accuracy-Flip: 0.50000+-0.00000 Training: 2021-09-03 04:06:46,999-[lfw][2000]Accuracy-Highest: 0.50000 testing verification.. (14000, 128) infer time 51.230545000000006 Training: 2021-09-03 04:07:42,615-[cfp_fp][2000]XNorm: 9.465774 Training: 2021-09-03 04:07:42,616-[cfp_fp][2000]Accuracy-Flip: 0.49986+-0.00043 Training: 2021-09-03 04:07:42,616-[cfp_fp][2000]Accuracy-Highest: 0.49986 testing verification.. (12000, 128) infer time 43.916425999999944 Training: 2021-09-03 04:08:30,471-[agedb_30][2000]XNorm: 9.467841 Training: 2021-09-03 04:08:30,471-[agedb_30][2000]Accuracy-Flip: 0.50033+-0.00067 Training: 2021-09-03 04:08:30,471-[agedb_30][2000]Accuracy-Highest: 0.50033

anxiangsir commented 3 years ago

maybe arcface causes some convergence problems, you can try cosface.

hhhThomas commented 3 years ago

Hi @anxiangsir , Thanks for your reply. I have tried out cosface and the test verification works correctly. However, I think it is not because of convergence reasons since the loss has decreased smoothly (53->2). Accuracy=0.5 was just like predicting all True or all False. Do you think resume training matters? I believe I have loaded the weights successfully since the loss value was preserved.