have people use inception-resnet-v2 to run the code?

[neklom]

You just change the argument of --model_def to point to the inception_resnet_v2

[hsm4703]

you say tend to run the code using inception-resnet-v2, with train_tripletloss.py you use your image to train can tell me How many classifications and How many pictures are in one category After the final training your inception-resnet-v2, with train_tripletloss.py how about performance better than inception-resnet-v1 ?

[hsm4703]

ok i will try it thanks

neklom notifications@github.com 於 2019年7月24日 週三 下午10:06寫道：

You just change the argument of --model_def to point to the inception_resnet_v2

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/davidsandberg/facenet/issues/1057?email_source=notifications&email_token=AKH2H5RB52QPQYG7ZCC3M6LQBBOXXA5CNFSM4IGFE6V2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2WOFRY#issuecomment-514646727, or mute the thread https://github.com/notifications/unsubscribe-auth/AKH2H5UK6ONBIFVWV3GPJZDQBBOXXANCNFSM4IGFE6VQ .

[neklom]

I don't really understand what you're trying to say I'm sorry, but I think you are referring to training using the triplet loss function, I tried using the tripletloss script and the inception_resnet_v1 architecture but I didn't have a good validation accuracy ! After 350 epochs I had like 86% on the training but nearly 16% on the validation rate ! for the images, I used the Casia-webface dataset.

[hsm4703]

I thought you used the database you created.
After 350 epochs I had like 86% on the training but nearly 16% on the validation rate !
is not good on accuracy
do you try add epochs 350 to 400 or more