search

astorfi / 3D-convolutional-speaker-recognition

:speaker: Deep Learning & 3D Convolutional Neural Networks for Speaker Verification
Apache License 2.0
780 stars 274 forks source link

minibatch loss not change #30

Closed Achillesheeel closed 6 years ago

Achillesheeel commented 6 years ago

I have use some utterances to test the code, and the program is running well, but the console prints that minibatch loss always during 7~9 and not decrease, accuracy=0, what's wrong about this? Thx!

astorfi commented 6 years ago

What dataset are you using for training?

Achillesheeel commented 6 years ago

I choose some utterances from internet about 1045 persons ,including 3000 utt, and I want to know which batchsize and epoch do I use to make the better results and let the loss minimize. Thx!

astorfi commented 6 years ago

Unfortunately, this question is vague to me. There is n magic number for that. This is base of trial an error. Start with 30 epochs for instance.