Low Accuracy and weird Triplet-Loss pattern when traning DNN model

[yimeng...@gmail.com]

Dear all:

I'm a graduate student and I'm trying to train new DNN models on my own datasets (all Asian faces, 110,000 pics with 14,000 person, 5-20 pics for each person), but all models show poor results (AUC about 0.70 for both LFW or own dataset)

I tried alpha with  0.05/ 0.1/ 0.2/ 0.3

other parameters as default:

peoplePerBatch: 15

imagesPerPerson: 20

epochSize: 250

The Triplet Error didn't decrease at all during each epochs. It shows very little fluctuation no matter how many epochs completed. So I usually stoped training after 10 epochs for each model.

Could anyone give suggestions?  Is the problem with DNN training?

Thanks!