Understanding Tensorboard Graphs

TTS (Text-to-Speech)
#1

Hi! I got a few question regarding the graphs visualized on the Tensorboard:

  1. Why there is no Validation Loss graph but instead EvalStat graph
    visualized there? Since during training i provide both Training set and
    Validation set.

  2. Why there is Evaluation Stat graph there? since i never provide Testing set
    during the training?

    What is the purpose of having EvalStat graph and what can we know
    from that graph?

  3. Among loss_decoder, loss_postnet and stop_loss, which one
    should be really important and to be focused on ?

  4. How can i know that my current models is overfitting or not since no
    Validation graph provided??

#2

Don’t “overfit” to names. If you mentally replace the name “Validation” with “Evaluation”, you solve your problems. In short, check Eval stats for validation.

If you are looking for the final voice performance of the model, “loss_postnet” is the target. But for the best, they all should converge.

1 Like
#4

Anwyay, im quite confuse, why do we need both Eval and Train Audios generated such as on the Tensorboard?

Sory to ask but, can you explain about this?

About TestAudios, is it generated by picking up phonemes from random sentences/audios?

#5

Eval == Validation, Train == Training

If you check train.py you see test audios are generated by a certain list of sentences.