Add data augmentation - Githubissues

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition

MIT License

1.03k stars 272 forks source link

Add data augmentation #59

Closed joonson closed 3 years ago

joonson commented 4 years ago

Online data augmentation (RIR and MUSAN noise)
Should retain backward compatibility with existing models and loss functions

ShaneRun commented 3 years ago

@joonson Sounds good. By the way, is there any tricks you will do to get earlier convergence instead of running for several hundreds of epoch?

joonson commented 3 years ago

@ShaneRun The reason that the ResNetSE34L takes so long to converge is because it has so few parameters. The very thin network takes much longer to converge. If you use the regular ResNet34 for example, it should converge to the same performance in less than 100 epochs.