search

jefflai108 / ASSERT

JHU's system submission to the ASVspoof 2019 Challenge: Anti-Spoofing with Squeeze-Excitation and Residual neTworks (ASSERT).
MIT License
56 stars 32 forks source link

the problem of training model by SENet34 #8

Closed wuqiangch closed 4 years ago

wuqiangch commented 4 years ago

@jefflai108 On dev database, I set NUM_SPOOF_CLASS=2,only can get eer=0.869. I set NUM_SPOOF_CLASS=10,only can get err=0.797. Can you give some help ? Thanks!

jefflai108 commented 4 years ago

From what I remember, it's pretty easy to get 0% EER on the DEV set. These are my suggestion:

  1. Can you first load the pre-trained models and see if you can get the reported EERs? If not, your setup may be wrong in the first place.
  2. I am guessing your feature may be wrong. Check out the feature preparation part in the paper.
  3. Train a few more times (for each model) and analyze the results.
wuqiangch commented 4 years ago

@jefflai108 I used the pre-trainded model on PA dev database, and get EER_CM: 0.592592592593. But in your paper the eer is 0.575.

jefflai108 commented 4 years ago

you mentioned you use Librosa for feature extraction. That may be another source of variation.

wuqiangch commented 4 years ago

@jefflai108 No, I use your code to extract feature.It means kaldi.

wangtao2668129173 commented 4 years ago

@jefflai108 On dev database, I set NUM_SPOOF_CLASS=2,only can get eer=0.869. I set NUM_SPOOF_CLASS=10,only can get err=0.797. Can you give some help ? Thanks!

hi, on dev dataset and eval dataset,l set class=2,and get eer 0.96,2.29 respectively,l could see your code.and no find num_classes=10 pretrained model, thanks