deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
22.18k stars 5.28k forks source link

Incorrect fc weights in [Resnet34, Arcface, MS1MV3] pretrained models #2056

Open zzzzz314314 opened 2 years ago

zzzzz314314 commented 2 years ago

Hello, for the [Resnet34, Arcface, MS1MV3], the fc weights might not be correct. The number of classes in the fc is inconsistent to the number of classes in MS1MV3 (91180!=93431) in both baidu and one drive.

https://pan.baidu.com/s/1CL-l4zWqsI1oDuEEYVhj-g#list/path=%2Fsharelink3494273403-838107067808247%2Farcface_torch%2Fms1mv3_arcface_r34_fp16&parentPath=%2Fsharelink3494273403-838107067808247

https://onedrive.live.com/?authkey=%21AFZjr283nwZHqbA&id=4A83B6B633B029CC%215579&cid=4A83B6B633B029CC

Would it be possible that you re-upload them? Many thanks

nttstar commented 1 year ago

the softmax fc weights should not be saved in our pretrained models.

thinh-re commented 1 year ago

Hi, thank you for your amazing work. I would like to have the Softmax FC weights for incremental learning experiment.

To be more clearer, @zzzzz314314 doesn't mean the softmax FC weights in pretrained model backbone.pt. He means that softmax FC weights in all the files rank_0_softmax_weight.pt, rank_1_softmax_weight.pt, ..., rank_7_softmax_weight.pt combine together, which is 91180 classes. But MS1MV3 dataset has totally 93431 classes.

ElisonSherton commented 9 months ago

Hi @zzzzz314314

Did you get any update about this? I also looked at the resnet50 versions and over there too, the softmax fc weights combined together leads to only 91180 classes instead of 93431 classes.

Alternatively, @nttstar , is there a dataset of 91180 classes for ms1mv3 which was cleaned further by insightface which we are not aware of; or which hasn't been publicly provided?

Thanks & Regards, Vinayak Nayak.

supertx commented 1 month ago

any upload message ? asking from 2024/6/3