search

philipperemy / deep-speaker

Deep Speaker: an End-to-End Neural Speaker Embedding System.
MIT License
905 stars 241 forks source link

Is the training process normal? #79

Closed wenku1992abc closed 3 years ago

wenku1992abc commented 3 years ago

image

is it overfitting? the train dataset loss decrease quickly,but the val_loss even no change. as shown in the picture @philipperemy

wenku1992abc commented 3 years ago

image then,early stopping itself.

philipperemy commented 3 years ago

@wenku1992abc yes it looks like you overfitted a bit your data. Your accuracy is 99.18% and your validation accuracy is around 90%. That's why the early stopping kicked in. The model is very big so it requires a lot of data. However, I guess your training is completed. An accuracy of 90% or above can be acceptable if you train from scratch of many many speakers. If you have more data, use it!

wenku1992abc commented 3 years ago

thanks your reply！ by the way 1080 speaker is it less? @philipperemy

philipperemy commented 3 years ago

LibriSpeech had 2484 speakers.

wenku1992abc commented 3 years ago

ok,thanks !