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

Mean and standard deviation comes NAN #14

Closed maulikjethva closed 6 years ago

maulikjethva commented 6 years ago

Hi Astorfi,

Your paper is awesome. I am trying to train speech data using 3D CNN. I have prepared data according to mention in paper. but during development phase I am getting mean and standard deviation "nan" in each epoch. I am getting following output:

Epoch 1, Minibatch 1 of 3 , Minibatch Loss= 2.1972, TRAIN ACCURACY= 0.000 Epoch 1, Minibatch 2 of 3 , Minibatch Loss= 2.2215, TRAIN ACCURACY= 0.000 Epoch 1, Minibatch 3 of 3 , Minibatch Loss= 2.2637, TRAIN ACCURACY= 0.000 TESTING after finishing the training on: epoch 1 Test Accuracy 1, Mean= nan, std= nan

Can you please help me, why am I getting this problem?

astorfi commented 6 years ago

@maulikjethva Thank you for your kind words ... The mean and std are simply the mean and standard deviation of the accuracy per batch. Please make sure that the accuracy calculations are correct. Can you tell me what is the value of test_accuracy_vector?