acc is always about 0.5 using mobilenetface

[tianxingyzxq]

testing verification..
(12000, 128)
infer time 30.116359
[lfw][6000]XNorm: 38.367005
[lfw][6000]Accuracy-Flip: 0.50000+-0.00000
testing verification..
(14000, 128)
infer time 35.065952
[cfp_fp][6000]XNorm: 38.365932
[cfp_fp][6000]Accuracy-Flip: 0.50000+-0.00000
testing verification..
(12000, 128)
infer time 30.366434
[agedb_30][6000]XNorm: 38.366582
[agedb_30][6000]Accuracy-Flip: 0.50000+-0.00000
[6000]Accuracy-Highest: 0.51533

the train script is
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network y1 --loss-type 4 --margin-s 128 --margin-m 0.5 --per-batch-size 128 --emb-size 128 --data-dir ../datasets/faces_ms1m_112x112 --wd 0.00004 --fc7-wd-mult 10.0 --prefix ../model-mobilefacenet-128

[nttstar]

I have no GPU server to test it right now. The author told me that he did the experiments by fine-tuning(train softmax firstly and fine-tuning with arcface loss).

[tianxingyzxq]

To pursue ultimate performance, MobileFaceNet, MobileFaceNet (112 × 96), and MobileFaceNet (96 × 96) are further trained by ArcFace loss on the cleaned training set of MS-Celeb-1M database [5] with 3.8M images from 85K subjects.
So the further train is fineturing ? not from scratch？

[tianxingyzxq]

I use ms1m train the mobilenetface network with softmax, the first testing verification is still 0.5
INFO:root:Epoch[0] Batch [1980] Speed: 785.83 samples/sec acc=0.004492
lr-batch-epoch: 0.1 1999 0
testing verification..
(12000, 128)
infer time 5.902649
[lfw][2000]XNorm: 3.298171
[lfw][2000]Accuracy-Flip: 0.50000+-0.00000
testing verification..
(14000, 128)
infer time 6.765173
[cfp_fp][2000]XNorm: 3.160100
[cfp_fp][2000]Accuracy-Flip: 0.50000+-0.00000
testing verification..
(12000, 128)
infer time 5.85961
[agedb_30][2000]XNorm: 3.269653
[agedb_30][2000]Accuracy-Flip: 0.50000+-0.00000
[2000]Accuracy-Highest: 0.50000

[marcosly]

@tianxingyzxq I face the same problem.However,i got different after several hours.
lr change to 0.001
lr-batch-epoch: 0.001 6241 18
testing verification..
(12000, 128)
infer time 8.014871
[lfw][140000]XNorm: 33.146315
[lfw][140000]Accuracy-Flip: 0.98867+-0.00552
testing verification..
(14000, 128)
infer time 9.476789
[cfp_fp][140000]XNorm: 29.224738
[cfp_fp][140000]Accuracy-Flip: 0.84671+-0.02232
testing verification..
(12000, 128)
infer time 8.189307
[agedb_30][140000]XNorm: 33.785845
[agedb_30][140000]Accuracy-Flip: 0.88883+-0.02323

[visionxyz]

The same with you, always 0.5.

[lmmcc]

i am also puzzled about the result...my result is like this,,,the train phase is reaching 10epcho,,,and the training acc is still 0.the lfw result is not as other researchers'. the highest is reaching to 0.79...

[nttstar]

I will provide a pretrained model soon.

[visionxyz]

@lmmcc I face the same problem when training with resnet-101. acc is low but lfw is good.

[wsx276166228]

@lmmcc The same problem with you! lfw accuracy=0.991,but train accuracy=0.000000. I wonder to know whether you solver the issue?

[wsx276166228]

@nttstar when I trained mobilefacenet with two 1080ti GPUs, and set batch_size to 256 per gpu. Through 20000 batchs, the lfw accuracy=0.991,but train accuracy=0.00000. I don't know if this phenomenon is normal？ If it is a problem, do you know what caused this phenomenon? Thanks!

[Wisgon]

@wsx276166228 What's the final accuracy of lfw,cfp_fp,agedb_30 you finally got?

[saxenauts]

@nttstar Hi, did you get a chance to upload the pretrained model for MobileFaceNet? Similar Problem here. Thanks.

[wsx276166228]

@Wisgon lfw=0.985 cfp_fp=0.854 agedb_30=0.921

[pribadihcr]

+1

[Wisgon]

I have change the learning rate from 0.1(by default) to 0.01, but the acc is still 0.And I'm still running to see the final result.

[wsx276166228]

@ShiyangZhang what's you mean? what's the learing rate you set? Thanks!

[tp-nan]

@wsx276166228 I forgot learning rate decay according to the original paper. Now I got a much better accuracy of lfw,cfp_fp,agedb_30. But it is not good enough,maybe my batch size(364) is too small. I'll report the accuracy latter.

[chenkingwen]

@wsx276166228 The same problem, do you have solved this problem?

[chenhuan19871014]

The same with you, always near 0.5.anyone knows how to solved it?

[staceycy]

Same problem. Anyone got any idea?

[baishiruyue]

@tianxingyzxq @nttstar i face the same question, the three valid data is always 0.5，and i check the params file of the model, i found that some layer`s paramters is near to zero，so the model can not be trained any more. how did you sovle this question?? thanks