Deepspeech model

[JRMeyer]

>>> sanjay.pandey
[September 20, 2019, 11:00am]

Hope you are doing well. slash
I just want to know deepspeech model 0.4.1 was trained on how many
epochs and how many hours of data was used and what was the final loss
of the model? slash
I am currently training on 130 hours of data from pretrained model i
should take the model till how much loss to consider it as
good?Currently i am training it for 100 epochs and it is running on 15th
epochs as of now with loss of 19. Thank you all in advance and thanks a
lot for always helping it wouldnt have been possible to come till where
we are without help of this forum and without all of you.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/deepspeech-model]

[JRMeyer]

>>> lissyx
[September 20, 2019, 11:21am]



> I just want to know deepspeech model 0.4.1 was trained on how many
> epochs and how many hours of data was used and what was the final loss
> of the model?

Please refer to the release page for v0.4.1 for those details.



> I am currently training on 130 hours of data from pretrained model i
> should take the model till how much loss to consider it as
> good?Currently i am training it for 100 epochs and it is running on
> 15th epochs as of now with loss of 19

You should also look at WER.

[Archived Post]

[JRMeyer]

>>> sanjay.pandey
[September 24, 2019, 6:25am]

for your reference.I
have taken all the batches train,test,dev as 1 and epoch as -100. i was
getting 23 loss till 30 epoch and now i am getting continuously inf as
loss. what is the problem in my approach?

[Archived Post]

[JRMeyer]

>>> lissyx
[September 24, 2019, 7:51am]



> what is the problem in my approach?

Without more context it's hard to tell for sure. But batch size of 1 is
not going to get you anywhere for sure.

[Archived Post]

[JRMeyer]

>>> sanjay.pandey
[September 24, 2019, 8:48am]

I stopped my training as it was continuously giving inf and after that
increased batch size of train,test,dev as 20,48,48 with epoch -10 and
tried training again it was giving inf on training and validation loss
at the first epoch itself and after that i increased my learning rate
from 0.0001 to 0.001 even after that i am receiving inf at the first
epoch itself. After going through forum i think so there might me
problem of corrupt file.

[Archived Post]