Tts: Tensorboard example outputs.

[erogol]

Below is how a healty training looks like. I put it here as a reference for people using TTS. This is a training on LJSpeech with default model configuration.

[IveJ]

How about your input data

[erogol]

@IveJ what do you like to know ?

[IveJ]

Hi Eren,

Could we minimize noise by tuning loss functions or another one, as
possible!?

Thanks.

On Wed, Jan 15, 2020, 07:48 Eren Gölge [e-mail] wrote:

@IveJ https://github.com/IveJ what do you like to know ?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/mozilla/TTS/issues/337?email_source=notifications&email_token=AEYAMLZSY7PUI43PTXVMXE3Q5ZMN3A5CNFSM4KGR7SR2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEI6VCTA#issuecomment-574443852,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AEYAML2ZUVPAY6TYWCS727LQ5ZMN3ANCNFSM4KGR7SRQ
.

[erogol]

All of these possible but I can't really comment without knowing all your data and training. Please post it on discourse.

[vcjob]

Hi @erogol
according to the plots, what's the right step to stop training? Is it still learning at 300k+ step? what's the main plot we are looking at to decide when to stop? I mean, it seems like on the stop_loss plot it does not learn any longer, the training is going on though. Wouldn't it hurt the final model performance?

[erogol]

@vcjob https://github.com/mozilla/TTS/wiki/FAQ#how-do-i-know-when-to-stop-training

[WeberJulian]

Hi, I have a similar looking tensorboard but sometime on the ground truth, I have a weird artefact that is maybe just the silence trim but the models is struggling to replicate the green rectangle. Is this normal behaviour ? Shouldn't it be just deep blue if it's silence ? Can it hurt performance ?

EDIT : If anyone has the same question @erogol responded here

[heoun]

@erogol - I'm running into an issue with tensorboard where "No scalar data was found." even though the output is the same as the one in the config.json. Any idea on how to resolve this issue? I have already ran the model and wanted to confirm the performance of my model via tensorboard.