I would like to know what validation loss did you guys receive at the end of 30 epochs for your pretrained model and also on how many hour data is this pre trained model trained on?
tuttlebr
(Tuttlebr)
April 19, 2019, 12:06am
2
This was posted by one of the Mozillians in March '18. See kdavis post.
We are trying to train Deepspeech model on our custom English data ~ 1000 hours. Unfortunately after 6 epochs val loss stops going down and it starts going up. Finally stopped after 9 epochs with not very good accuracy of the 6th epoch model. Data label quality is very good.
The hyper parameters used were exactly same as the deepspeech release version.
One thing we suspect is that the hyper parameters need to be changed and need to experiment with different values. But just wondering how to go…
The number of training data hours isnt mentioned there.
And why are there only 12 epochs, the model was trained for 30 epochs right?