Performance and backbone network

[John1231983]

Thanks for sharing a great work. In your code, you provided some backbone network

support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152']

In the Model Zoo , the insightface provide LResNet100E-IR

Method	LFW(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace(%)
Ours	99.77	98.27	98.28	98.47

This ís your result that trained from scratch with IR-101 using your setting.

Evaluation: LFW Acc: 0.9976666666666667, CFP_FP Acc: 0.983, AgeDB Acc: 0.9818333333333333, CPLFW Acc: 0.9301666666666666 CALFW Acc: 0.9608333333333332

And this is your report

Data	LFW	CFP-FP	CPLFW	AGEDB	CALFW	IJBB (TPR@FAR=1e-4)	IJBC (TPR@FAR=1e-4)
Result	99.80	98.36	93.13	98.37	96.05	94.86	96.15

I have some questions after reading your code:

1. Do you use augmentation (I found only RandomHorizontalFlip has been applied)? Insightface team used augmentation such as flip, ColorJitterAug, compress_aug... https://github.com/deepinsight/insightface/blob/3866cd77a6896c934b51ed39e9651b791d78bb57/recognition/image_iter.py#L207?

2. I am using 4GPU with batch size of 700/each GPU. My performance is smaller than your report. Do you think number of GPU is the reason (you used 8 GPUs)?

3. Does your IR_101 same with LResNet100E-IR in term of number of FLOP and params? I found that you save backbone and head seperately, while insighface saved them into one model? Any difference?

4. Have you measure the inference speed of IR_101? I feel it too slow than mxnet

[HeshamAMH]

please John1231983, I test the model provided by the author on LFW but the result is 0.661 not 99.+. So, please I need a link to LFW and its pair file. I think the problem for me is LFW itself.

Thanks in advance

[quangtn266]

please John1231983, I test the model provided by the author on LFW but the result is 0.661 not 99.+. So, please I need a link to LFW and its pair file. I think the problem for me is LFW itself.

Thanks in advance

Hi, Did you solve your problem?

I also have the same problem and want to know any solution for that issue.

[marcohuber]

@quangtn266
I downloaded the (preprocessed) LFW data including the pairs file from here and achieved at least on LFW = 99.783 using the provided CurricularFace model. Still a bit worse, but I guess it's the minor differences in preprocessing. Maybe something is wrong with your preprocessing/pairs.

[michellerybak]

I had this problem when I didn't specify the checkpoint in the config file under BACKBONE_RESUME_ROOT