Error occurred while training the FullSubNet-Plus model

[Yaho-b]

Hi,
I encountered an error like that when training FullSubNet-Plus about 100 epoch on DNS dataset. What is the reason and how to solve it.

error:
File "/venv/py365/lib/python3.6/site-packages/librosa/util/utils.py", line 310, in valid_audio
raise ParameterError("Audio buffer is not finite everywhere")
librosa.util.exceptions.ParameterError: Audio buffer is not finite everywhere.

Looking forward to your reply!!

[RookieJunChen]

When the model is being trained, there may be problems with the parameters that cause the final output audio to have nan values in the numpy vector, causing librosa to fail to write the numpy vector to a wav file. This may happen depending on the probability, you can continue to retrain or choose the best epoch so far to use.

[Yaho-b]

OK, Thanks! By the way, what number should max_epoch be set? the orig train.toml set epochs=9999.... Maybe it is too big~~

[RookieJunChen]

That value is just an arbitrary setting I made.
You need to judge for yourself whether the model has converged or not.

[mystlee]

That value is just an arbitrary setting I made. You need to judge for yourself whether the model has converged or not.

Hi,
Can you tell me when it was converged in terms of epoch?