search

davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.8k stars 4.81k forks source link

Accuracy reaching 100% #956

Open taureanamir opened 5 years ago

taureanamir commented 5 years ago

@davidsandberg I am training on my own dataset, 92 classes approx. 15 images per class, using softmax loss but the training accuracy is reaching 100%. this is very unusual. There must be something very fishy going on. Can you point me to the right direction?

luameows commented 5 years ago

I got the same problem as u did. While i trained webface dataset, the training acc still became 1. At first, i thought it was my dataset not balanced. Then, i did random flip for training set, and added dropout layer at the last fully connected layer, the situation didn't change.

taureanamir commented 5 years ago

@luameows Do you know if anyone else faced the similar issue??

luameows commented 5 years ago

@taureanamir As far as i know, ur dataset is too small for this deep network. Although softmax is not very strict for classification or verification, only 92*15 images can still cause overfit. Try to change a light CNN network or get more images for ur training may solver this situation.

taureanamir commented 5 years ago

@luameows Thank you very much for your suggestion.