A healthy Tensorboard log for Tacotron.

[erogol]

Below is how a healthy training looks like with Tacotron. I put it here as a reference for people using TTS.

This is Tacotron2 with LJspeech dataset and gradual training is enabled.

[thorstenMueller]

Hi @erogol .
Thanks for sharing tensorboard (best practice) graphs 👍 .

Maybe it might be helpful for beginners to tell that gradual training switches learning parameters based on a reached step count. This impacts the graphs shown here.

In this case (i guess) gradual switched values on:

• 50k steps
• 130k steps
• 280k steps

[erogol]

ignore (move the thread up again)