qidiso / mobilefacenet-v2

🔥improve the accuracy of mobilefacenet(insight face) reached 99.733 in the cfp-ff、 the 99.68+ in lfw,96.71+ in agedb30.🔥

mobilefacenet-V2

now we get more higher accuray:

[lfw][12000]Accuracy-Flip: 0.99667+-0.00358
[agedb_30][12000]Accuracy-Flip: 0.96667+-0.00167 use my modified mobilenet network.

lr-batch-epoch: 0.01 11738 1 testing verification.. (12000, 512) infer time 39.129495 [lfw][36000]XNorm: 22.729305 [lfw][36000]Accuracy-Flip: 0.99667+-0.00358

improve the accuracy of mobilefacenet in paper mobilefacenet论文(https://arxiv.org/abs/1804.07573)

First step training (use softmax to pretrain): train softmax(facenet):

[lfw][62000]XNorm: 23.029881 [lfw][62000]Accuracy-Flip: 0.99383+-0.00308 testing verification.. (14000, 512) infer time 20.121058 [cfp_fp][62000]XNorm: 24.043967 [cfp_fp][62000]Accuracy-Flip: 0.89343+-0.01705 testing verification.. (12000, 512) infer time 16.860138 [agedb_30][62000]XNorm: 23.566453 [agedb_30][62000]Accuracy-Flip: 0.93883+-0.01675 saving 31 INFO:root:Saved checkpoint to "../models/MF/model-y1-softmax12-0031.params"

pretrained models: https://pan.baidu.com/s/1xBq9FoL79z7K892aFWkmFw

Second step: CUDA_VISIBLE_DEVICES='0,1' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --margin-s [128] --lr-steps 120000,180000,210000,230000 --emb-size [512] --per-batch-size 150 --data-dir ../data/faces_ms1m_112x112 --pretrained ../models/MobileFaceNet/model-y1-softmax,20 --prefix ../models/MF/model-y1-arcface

Third step: CUDA_VISIBLE_DEVICES='0,1' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.001 --lr-steps 40000,60000,70000 --wd 0.00004 --fc7-wd-mult 10 --emb-size 512 --per-batch-size 150 --margin-s 64 --data-dir ../data/faces_ms1m_112x112 --pretrained ../models/MF/model-y1-arcface,46 --prefix ../models/MF/model-y1-arcface

Update wd=0.00001 , --fc7-wd-mult 10 --emb-size 512 i get new Accuracy:

Accuracy

dbname	accuracy
lfw	0.996233
cfp_fp	0.94300
age_db30	0.96383

##########first #CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.1 --emb-size 512 --per-batch-size 240 --margin-s 64 --wd 0.00001 --fc7-wd-mult 10 --data-dir /Users/sunyimac/faces_emore --pretrained ../models/MobileFaceNet/model-y1-arcfaced,18 --prefix ../models/MobileFaceNet/model-y1-arcface

#CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.01 --emb-size 512 --per-batch-size 240 --margin-s 64 --wd 0.00001 --fc7-wd-mult 10 --data-dir /Users/sunyimac/faces_emore --pretrained ../models/MobileFaceNet/model-y1-arcface,62 --prefix ../models/MobileFaceNet/model-y1-arcfaced

CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.00001 --emb-size 512 --per-batch-size 240 --wd 0.00001 --fc7-wd-mult 10 --data-dir /Users/sunyimac/faces_emore --pretrained ../models/MobileFaceNet/model-y1-arcface,75 --prefix ../models/MobileFaceNet/model-y1-arcfaced

Update wd=0.000001 trainning is not end. now is the new Accuracy: i get new higher Accuracy:

Accuracy

dbname	accuracy
lfw	0.99667
cfp_fp	0.94300
age_db30	0.96700

Update wd=0.0000001 trainning is not end. now is the new Accuracy: i get new higher Accuracy:

Accuracy🔥

dbname	accuracy
lfw	0.99683
cfp_ff	0.99733
cfp_fp	0.94500
age_db30	0.96717
you can visit my log file:
https://github.com/qidiso/mobilefacenet-V2/blob/master/retrain0.001.log

Now Release the models:

[models:]https://github.com/aidlearning/AidLearning-FrameWork/tree/master/src/facencnn/models (reached 99.733 in the cfp-ff、 the 99.68+ in lfw,96.71+ in agedb30)

mobilefacenet-v2's Issues

Problem with the module symbol_utils

I suggest you add symbol_utils.py to your code. Thank you so much.

for evaluate code

Thanks for your great work! Could you share me the code to evaluate the accuracy with my model for dataset LFW，CFP and AgeDB! .Thanks a million.

Can you tell us something about what improvements you have been made?

As the title , can you tell me the difference between yours and mobilefacenet?
Some details，such as LeakyReLU and wd

extremely grateful |

i use your pretrained model to test it seems could't work.

whatever the pair face image that I input ,the cosine sim always in around 0.60.

pretrained models

Could you please upload your highest accuracy pretrained models? @qidiso

The reason why the accuracy is so high

Hello, you can train such a good model, what parameters do you mainly adjust, can you share your experience?

Model Size 180 MB?

You have suggested three steps of training rght?, for the first step, what should be the expected model size? I am getting 180 MB, but you final model is 5 MB.
I keyed in the command you provided in the README. Is 180 MB expected or am I doing anything wrong?

Resume Training?

Hi, how do you resume the training with the last saved checkpoint after the training was stopped?

Can you provide the network structure file of mobilefacenet-v2? Thanks

Question about improve

It seems that you only change output dimension from 128 to 512?

请问这个模型怎么移植到安卓手机上去呢

Incompatible attr in node when train a pretrained model. any help would be appreciate

MXNetError: [13:28:27] /home/travis/build/dmlc/mxnet-distro/mxnet-build/3rdparty/mshadow/../../src/operator/tensor/../elemwise_op_common.h:135: Check failed: assign(&dattr, vec.at(i)): Incompatible attr in node at 0-th output: expected [85164,512], got [85742,512]

Why is acc always zero

I used train_softmax,py in https://github.com/deepinsight/insightface/blob/master/src/train_softmax.py
The parameters are set like this：
CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.00001 --emb-size 512 --per-batch-size 80 --wd 0.0000001 --data-dir /home/mi5/insightface/src/data/faces_vgg --pretrained /home/mi5/insightface/models/model-y1/model,00 --prefix ../models/MobileFaceNet/model-y1
Why is acc always zero？Can you help me?

Pretrained Model Confusion...

I was looking at the models repo and was confused. What is the difference between det1, det2, and det3 as I am trying to use caffe and are they all for mobilefacenet V2.

Can you share parameter file (.params)? I cant load pretrained model (-symbol.json) without it.

share the inference file to Visualization the results?

I want to detect my custom face with name on the result window with a pre-trained model of mobilefacenets

What is the "loss-type==4"?

The project insightface has changed so much. What is the loss corresponding to this loss type?

where is the file train_softmax.py？？？？？

emb-size

why your emb size is 512? the result is unfair compared to the orginal paper.and the model size is too large

why the pretrained model with low accuracy

Why I use your training model with low initial accuracy? Am I wrong?

cmd:CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.01 --lr-steps 40000,60000,70000 --wd 0.00004 --fc7-wd-mult 10 --emb-size 512 --per-batch-size 90 --margin-s 128 --data-dir ../datasets/faces_ms1m_112x112 --pretrained ../models/MF/model-y1-softmax12,31 --prefix ../models/MF/model-y1-arcface >>& file.txt &

first test:
testing verification..
(12000, 512)
infer time 154.665504
[lfw][2000]XNorm: 22.648476
[lfw][2000]Accuracy-Flip: 0.98717+-0.00563
testing verification..
(14000, 512)
infer time 175.866512
[cfp_fp][2000]XNorm: 19.110292
[cfp_fp][2000]Accuracy-Flip: 0.83271+-0.01969
testing verification..
(12000, 512)
infer time 157.715533
[agedb_30][2000]XNorm: 22.118053
[agedb_30][2000]Accuracy-Flip: 0.90683+-0.01981
saving 1
INFO:root:Saved checkpoint to "../models/MF/model-y1-arcface-0001.params"

Could you give me the link to the test dataset

Thank you very much for your work. Could you give me the link to the test dataset,such as lfw,agedb_30......

how to find the peretrain models

how to find the peretrain models,i cannot find in this:
https://github.com/aidlearning/AidLearning-FrameWork/tree/master/examples/facencnn
cannot find weight files?

Why is the modification of the network model bottleneck so different from the original text? Is there any basis?

thnx you best work!!
Why is the modification of the network model bottleneck so different from the original text? Is there any basis?some basis is importance.

Model size

I see https://github.com/deepinsight/insightface/wiki/Model-Zoo the MobileFaceNet size only 4.1M, This model is too large