search

astorfi / 3D-convolutional-speaker-recognition

:speaker: Deep Learning & 3D Convolutional Neural Networks for Speaker Verification
Apache License 2.0
778 stars 275 forks source link

loss = 0,train acc = 0 #25

Closed duynguyen5896 closed 6 years ago

duynguyen5896 commented 6 years ago

Hi astorfi, I'm trying to use your code to train a model with 31 labels, 60 samples for each label. However, when i use train_softmax.py, last minibatches return loss = 0 while train acc = 0. Do you have any idea to fix it?

image

Thank you.

duynguyen5896 commented 6 years ago

Hi astorfi, i see you have "train_softmax.py" for training model. Is this file just used for training all layers of model?

astorfi commented 6 years ago

@duynguyen5896 train_softmax.py is just for training all the layers as you mentioned. It is created for generating the background model. It is hard to say why it's not converging because I do not have access to your data or the details.

duynguyen5896 commented 6 years ago

@astorfi, for the first problem, i realized that my label was higher than 31 and after relabeled from 0 to 30, it works. Thank for your answer, it helps me a lot.

astorfi commented 6 years ago

@duynguyen5896 Good to hear that.

duynguyen5896 commented 6 years ago

@astorfi , can you tell me how many samples per person you used to train and validate your background model?

astorfi commented 6 years ago

Please refer to this issue for futher details.

duynguyen5896 commented 6 years ago

Thank for your help again @astorfi !!!

astorfi commented 6 years ago

@duynguyen5896 My pleasure