If trained with MS-Celeb-1M or VggFace2 dataset ,how much can the test result be improved?

[LingJiangXie]

[MuggleWang]

I don't know because I have never done such an experiment. You can try it with a simple modification.

[gdshen]

I have trained on the msceleb cleaned by insightface, and have achieved 99.6~99.7 accuracy in LFW with resnet50.

[Sayyam-Jain]

I have trained on the msceleb cleaned by insightface, and have achieved 99.6~99.7 accuracy in LFW with resnet50.

Can you please share the model weights?

[akumar14]

Can you also share the model on One Drive ?

[weihua04]

I have trained on the msceleb cleaned by insightface, and have achieved 99.6~99.7 accuracy in LFW with resnet50. @gdshen i use the msceleb cleaned by insightface 、 LResnetE50-IR + cosFace ,but the entropy loss is 10.9 and didn't decline, 93%~94% accuracy in LFW , have u ever met this problem

[MuggleWang]

Trained on VggFace2 with Sphere20. Achieved 99.55% on LFW. (Note: Input is 112X112)

• Baidu Disk 链接: https://pan.baidu.com/s/16kWW4Gq6TdPRn2rA2ickkQ 提取码: pya3

• Google Drive https://drive.google.com/file/d/19RAVxfvw15tavjpcktBBxFygxxTFM7Z1/view?usp=sharing

[HeshamAMH]

please MuggleWang when I test your model (net_sphere20_data_vggface2_acc_9955.pth), I obtained an error:

RuntimeError: Error(s) in loading state_dict for sphere: size mismatch for fc.weight: copying a param with shape torch.Size([512, 25088]) from checkpoint, the shape in current model is torch.Size([512, 21504]).

Thanks in advance

[MuggleWang]

@HeshamMM Check your input size (112x112), and accordingly the last fully connected layer in the model file layer.py should be self.fc = nn.Linear(512 * 7 * 7, 512). And if input size is 112x96, it should be self.fc = nn.Linear(512 * 7 * 6, 512). Further, in the inference(test) stage, match them with the model *.pth.

[HeshamAMH]

Deeply thanks Dr. MuggleWang.