davidsandberg
poor performance on Asian faces
I'm try to compute the cosine similarity of a couple of Asian faces, but it seems facenet can be confused by these faces (two total different faces can have 0.7 or higher cosine similarity). Is this normal or I am missing something ?
If I train facenet on a Asian Face dataset, will it help this problem?
Currently I only have a gtx1050 at hand, will it be sufficient to train the facenet?
Thanks
I think you can use the Asian Face dataset to fine-tune the models, and the result will be better.
I'm also working on asian face recently，have you solved the problem by training on asian dataset，or solved it in other ways? thank you very much，present poor accuracy worries me
@0liliulei I didn't use asian dataset to train the model for the lack of asian data, I just use about 400 classes asian dataset to fine-tune the model provided by Davidsandberg. I didn't do many tests yet, but it seems that the new model's performance is better than the old one on asian faces. I think if you have enough asian data and you can train the model by your dataset, or just fine-tune the premodel. The dataset really has a great influence on the model. I hope that if you get some great results, please let me know.
@angushh Thank you for your prompt reply.I am now applying for an Asian face dataset.If I get substantial progress through this dataset,I will @ you in time.
@0liliulei how to fine tune the faceNet ,could you share the way for me.Thank you very much.
We are also trying to fine-tunning the pretrained model on a asian face dataset, maybe we can share the dataset to get more progress on this matter? Sounds good?
@yanmenglu just to continue training on the model davidsandberg provided with asian dataset
@0liliulei I also want to train on asian dataset, but I have no idea, can you share your idea?Thanks!
I think the true issue is the quality and quantity of the dataset, which means the dataset should contains enough images (millions) and enough variance(age, pose, lighting etc..). The Performance of LFW is just a start, and far from real application. This paper really reveal some interesting issue about the LFW benchmark.
@angushh @0liliulei Hi,I have also used about 7000 classes(23500 imgs) asian dataset to fine-tune the pretrained model(MS-celeb-1M dataset) provided by Davidsandberg to improve the performence on asian faces.How about your lfw Accuracy? I got a low lfw accuracy about 93.7% after 15 epochs lower than David's pretrained model and the trend is not optimistic.May be something wrong?Here is some result below:
- command line arguments: --logs_base_dir ~/logs/pretrained_facenet/ --models_base_dir ~/models/pretrained/ --data_dir /home/linzai/PycharmProjects/facenet/data/mydataset160 --image_size 160 --model_def models.inception_resnet_v1 --lfw_dir /home/linzai/PycharmProjects/facenet/data/lfw_mtcnnpy_160 --optimizer RMSPROP --learning_rate -1 --max_nrof_epochs 160 --keep_probability 0.8 --random_crop --random_flip --learning_rate_schedule_file data/learning_rate_schedule_classifier_mydataset.txt --weight_decay 5e-5 --center_loss_factor 1e-2 --center_loss_alfa 0.9 --learning_rate_schedule_file /home/linzai/PycharmProjects/facenet/data/learning_rate_schedule_classifier_mydataset.txt --pretrained_model /home/linzai/PycharmProjects/facenet/src/20170512-110547/model-20170512-110547.ckpt-250000
2.fine-tune on learning rate and write in learning_rate_schedule_classifier_mydataset.txt:
0: 0.1 30: 0.01 50: 0.001 110: 0.0001
3.lfw_result.txt:
0 0.95467 0.77300 1000 0.94717 0.65133 2000 0.94700 0.70333 3000 0.94700 0.63900 4000 0.94317 0.64167 5000 0.94133 0.61067 6000 0.93900 0.58033 7000 0.93633 0.59700 8000 0.94183 0.62500 9000 0.93867 0.59000 10000 0.93683 0.59267 11000 0.93533 0.58333 12000 0.92650 0.49200 13000 0.91650 0.50700 14000 0.93567 0.61233 15000 0.93767 0.60233
Can someone give me some advice?THX!
@Linzai1994 I think,first,though your dataset and MS dataset are all people faces,but your dataset maybe have absolutely different quantity with MS,to be precise,your faces' styles are different from MS's.although MS have Asian faces,but most of them are Caucasian faces,so it can work well on lfw. second,as you have fine-turn the model with Asian dataset,this model will show preference to Asian face,and it will loss initial keen eye to Caucasian faces,because they have different styles.So I think you need to test or validate on Asian dataset. third,if after you test on Asian dataset,but it still performance not well.I just want to say it's normal,I think if we want to get a 99% accuracy model on any Asian dateset,we need billions of samples.otherwise,we should make sure that training set and test set are as similar as possible.what I mean is we just train a model which only can be use in a small scale. finally, lfw accuracy prove nothing if you want to build a model which can be put into practice.
@0liliulei Thank for your reply! 谢谢~ I see what you mean.The dataset determines what the model outputs.I used the model which fine-tuned by my dataset in real scene camera and the result is still not satisfactory.I think the comparison result also rely heavily on the train data bias, cross factors,e.g. angle,light,age difference and image quality.
@Linzai1994 Would you mind sharing your dataset or where you got it from? Thanks!
@harryxu-yscz Sorry,it's a private dataset from company .I’m afraid I can't share with you~
You can have one more solution, you can add images of people from your organization to large dataset like Ms dataset and train for complete set. In this way you will have accuracy for your dataset as well as for lfw dataset.
@tenggyut @angushh @0liliulei @yanmenglu @Zumbalamambo @Linzai1994 Hello everyone! I have same project with you guys, training model on Asian faces. Validation set is from customer(Asian faces). Use model provided by sandberg, the validation acc is about 75% at the begining. After finetuning on Asian faces, validation Acc on asian faces reaches 92%. And of course, the performance on lfw will reduced a little (not sure how much though)...
Training set : 10,322 ids, around 350,000 images, collected from Web Validation set : from customer
About how to keep performance on lfw while generalize better on other private set (like Asian faces), this paper may help: https://arxiv.org/abs/1606.09282
I'm currently woring on implementing this paper "Additive margin softmax(AM-softmax)": https://arxiv.org/abs/1801.05599 With all configuration same as author's, I only achieved 98.3% on lfw, only difference is author use caffe while I use tensorflow. In paper, author said Mom is better than Adam, but Mom for me only reaches 94.3% while Adam can easily reach 98%. My hypothesis is the difference of implementing details of framework(not so sure, but other parts are exactly same with author's). Would be glad if someone can discuss with me.
I have just started it. I have got one model trained for Microsoft 20K identities, and working on Microsoft 100K. Microsoft 100K dataset is divided into 20 equal parts. Each part I am training with microsoft 20K pretrained model. Currently only 2 parts are done. Once all parts are done, I will make the model available.
On Mar 17, 2018 12:38, "TaeWoo Kim" [email protected] wrote:
@Joker316701882 https://github.com/joker316701882
Are you releasing this model with source ?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/davidsandberg/facenet/issues/591#issuecomment-373899856, or mute the thread https://github.com/notifications/unsubscribe-auth/AFLq4mK95dyYxiFlbTdqdko_Tl0CVYLDks5tfLZdgaJpZM4RMidR .
Is there any public dataset specifically by race ? what are you guys using? (yes I know.. i can scrape.. but i'm curious if anyone has already built such)
@look4pritam look forward to seeing it. BTW.. what dataset is that? What's the URL for it?
Check for Microsoft 1 million challenge dataset. You will get 5 links 20GB 4 files and 4GB one file, which were released last year by Microsoft as two challenges.
On Mar 17, 2018 16:12, "TaeWoo Kim" [email protected] wrote:
@look4pritam https://github.com/look4pritam look forward to seeing it. BTW.. what dataset is that? What's the URL for it?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/davidsandberg/facenet/issues/591#issuecomment-373910684, or mute the thread https://github.com/notifications/unsubscribe-auth/AFLq4uNrixbRgLCQgNlkaMyM82bo_xBfks5tfOiMgaJpZM4RMidR .
Hi @Joker316701882 Did you train the model with softmax_loss or triplet_loss? If it is possible, can you provide the parameters of the training at that time?
Thank you in advance！
@taewookim @nellycui Sorry can't release model. But I do implemented training logic of 'Additive Margin Softmax (or CosFace by Tencent)'. It works extremely better than any loss function in this repo. Here is source code: https://github.com/Joker316701882/Additive-Margin-Softmax
Dear All, About fine-tuning, is there any particular strategy?
Do you update every trainable variables? My rough idea currently is, maybe just fine-tuning some higher level weights would be good enough? Maybe like tensorflow-for-poets?
https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0
My reasoning is that the first few layers of David's pretrained model is probably already good at recognizing basic visual elements such as oriented edges and opposing colors etc.
Maybe we only need to fine tune upper layers' weight?
Does anyone have any updates? Seems like a pretty hot topic
@look4pritam Any update on the release?
@angushh, thank you for your information, did you use the train_softmax.py and ckpt file, keep training to fine tune the pretrained model?
If so, did you face the problem that the output category of classifier was not match with Daivd’s retrained model?
Here’s the way I did, i just changed the size of “train_set “ to match original category of pretrained model(2017*****)
For example, i used 500 different people with 1000 images, so the output size of softmax FC layser is 500. However, because the original FC output is 4xxxx , so I need to forcely change my train_set size to 4xxxx to match the pretrained model.
Is that correct?
Thank you for your info.
