search

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1.03k stars 272 forks source link

I need not only test accuracy but test loss also... #45

Closed zeek-han closed 4 years ago

zeek-han commented 4 years ago

When i train with default setting, VEER is calculated for every 10 iteration, and I regard VEER as Validation Equal Error Rate. thus, VEER is derived from test accuracy and it is very important measure. But I need not only VEER but Validation Loss also to pick the best model. How could I print Validation loss?

Thank you...

joonson commented 4 years ago

There is no 'validation loss' as such, since the train function is not run on the validation set. To implement this function, you will need to:

  1. Initiate data loader for the validation set
  2. Modify train_network function to run with self.eval() mode.
  3. Run this function without the weight updates.
zeek-han commented 4 years ago

Oh, I think I express improper way thus you misunderstood my purpose... The sample code you give just train on training set, and for every 10 epoch, calculate EER on test set and print it. how could I print LOSS on TEST SET as well as equal error rate for every 10 epoch?

thank you...

joonson commented 4 years ago

I think my answer still remains the same. You will need to run the train function without the weight update every 10 epochs. Currently the script does not compute the validation loss.

zeek-han commented 4 years ago

Oh, thank you for your answer! I understand it!