search

TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.73k stars 423 forks source link

Did you reproduce megaface ? #29

Open twmht opened 5 years ago

twmht commented 5 years ago

Hi

Did you reproduce megaface result on pytorch?

thank you

TreB1eN commented 5 years ago

no, you can have a try

twmht commented 5 years ago

@TreB1eN

I tried mobilenet, but can't reproduce the paper's result (92%)

My result is only 87%

TreB1eN commented 5 years ago

I don't have clue about training Megaface, normally there are a lot of noises in the training set, how about try use emore dataset pretrained model here to test in Megaface test set ? See what is the accuracy ?

twmht commented 5 years ago

@TreB1eN

I did not train on megaface, I trained on refine MS-1M and test on megaface. Did you try to do that?

TreB1eN commented 5 years ago

I haven't

ruiming46zrm commented 5 years ago

@twmht

can you tell us your acc on cfp-fp by res50, mine is 87.5%

TreB1eN commented 5 years ago

0.9504 CFP-FP(%) by IR-SE50, pretrained model are provided

ruiming46zrm commented 5 years ago

0.9504 CFP-FP(%) by IR-SE50, pretrained model are provided

training refined-ms1m ,I can get LFW 99.52% but can't get cfp-fp so well , maybe something wrong.

twmht commented 5 years ago

@TreB1eN

Did you trained on refined-ms1m on IR-SE50?

twmht commented 5 years ago

@ruiming46zrm

Did you use this code without modification? I haven't tested cfp-fp yet.

TreB1eN commented 5 years ago

@twmht yes

ruiming46zrm commented 5 years ago

@twmht yes, ok

twmht commented 5 years ago

@TreB1eN

I can help to test the provided model in megaface.

ruiming46zrm commented 5 years ago

@twmht will you share mega test code if it's convenient~ ?

twmht commented 5 years ago

@ruiming46zrm

most of testing code is from insightface. Did you tried to reproduce the result on insightface?

ruiming46zrm commented 5 years ago

@twmht thanks, I don't know much about the megaface, I'm gonna try.

twmht commented 5 years ago

@TreB1eN and @ruiming46zrm

I just tried @TreB1eN 's provided resnet50 pretrained weight, but only got 90% on megaface.

but from deepinsight (https://github.com/deepinsight/insightface), they provide resnet50 which can get 97.64% on megaface.

since this is a huge gap, there might be some issues with pytorch implementation.

twmht commented 5 years ago

@TreB1eN

in addition, mobilenet only got 74% on megaface.

ruiming46zrm commented 5 years ago

provided resnet50 get 91.6% on megaface, my training only 89.9%

yangyuke001 commented 5 years ago

@twmht will you share mega test code if it's convenient~ ? mainly gen_megaface.py for pytorch version,thanks a lot

ruiming46zrm commented 5 years ago

@yangyuke001 I have one , what's your weichat?

yangyuke001 commented 5 years ago

yangyuke001

At 2019-01-07 08:23:09, "ruiming46zrm" notifications@github.com wrote:

@yangyuke001 I have one , what's your weichat?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

ruiming46zrm commented 5 years ago

yangyuke001 At 2019-01-07 08:23:09, "ruiming46zrm" notifications@github.com wrote: @yangyuke001 I have one , what's your weichat? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

user name unexist , mine is : ruiming46zrm

cvtower commented 5 years ago

Hello @ruiming46zrm, @twmht

The mega test code has to load a trained model and do inference, i mean it can not be tested during training, right? Then i have a question, this repo saves model per few iters, however, the best accuracy on different dataset may not always the same saved model. Shall we just take the best accuracy on each dataset as the final performance?

ghost commented 4 years ago

Hi, Do we need to normalize the images of Facescrub and Megaface before feeding them to the model to get features? Thank you for reading.

gaow0007 commented 4 years ago

@TreB1eN and @ruiming46zrm

I just tried @TreB1eN 's provided resnet50 pretrained weight, but only got 90% on megaface.

but from deepinsight (https://github.com/deepinsight/insightface), they provide resnet50 which can get 97.64% on megaface.

since this is a huge gap, there might be some issues with pytorch implementation.

Could you share the test configs of MegaFace? Such as select how many images as galleryset?

HulkMaker commented 4 years ago

@TreB1eN and @ruiming46zrm I just tried @TreB1eN 's provided resnet50 pretrained weight, but only got 90% on megaface. but from deepinsight (https://github.com/deepinsight/insightface), they provide resnet50 which can get 97.64% on megaface. since this is a huge gap, there might be some issues with pytorch implementation.

Could you share the test configs of MegaFace? Such as select how many images as galleryset?

Could you share Megaface dataset since its stopping supported in their website. And I would like to test on it too. Thanks!