Handle inf in deepspeech 0 5 1

[JRMeyer]

>>> vigneshgig
[August 23, 2019, 6:22am]

Hi, In new alpha version the mozilla deepspeech team included a code to
get the inf loss file.but, alpha version does not have pretrained model
and also if you have huge dataset you have to wait for whole epochs to
get the all inf loss file and also I find that if more inf file are
trained , the good file files also getting inf loss.So I added line of
code to convert the inf loss to zero so that the inf loss file will not
get trained on the model. and also you can get the mostly real bad inf
loss file. slash
Here the code I modified Deepspeech.py.

DeepSpeech_testing.py.zip
(10.4 KB) slash
and you have to just copy and paste alpha version feeding.py and text.py
in 0.5.1 ulti folder.

Thanks

[This is an archived TTS discussion thread from discourse.mozilla.org/t/handle-inf-in-deepspeech-0-5-1]

[JRMeyer]

>>> reuben
[August 23, 2019, 7:23am]

Your changes aren't doing what you think they're doing. Setting the loss
to zero doesn't make it so that 'inf loss file will not get trained'.
The model will still be trained the same way it was before. The correct
solution is to remove or fix problematic files so that they don't
produce inf loss.

[Archived Post]