About the training/eval loss

[yl4579]

About the training/eval loss

[yl4579]

I think that was a mistake. I mislabeled the F0 and silence loss for training. I have fixed it.
I believe 60 hours are enough, so you may not need to train for 100 epochs. This number of epochs is for data augmented audios (with noises and reverberations) for voice conversion under more realistic conditions.

[yl4579]

I don't have it now, but if you have trained your own model I can tell whether it looks right or not.

[Charlottecuc]

@yl4579

--- epoch 2 ---
train/loss : 4.4963
train/f0 : 0.1389
train/sil : 4.3574
train/learning_rate: 0.0003
eval/loss : 2.7822
eval/f0 : 2.6451
eval/sil : 0.1372

--- epoch 50 ---
train/loss : 1.2654
train/f0 : 0.1041
train/sil : 1.1613
train/learning_rate: 0.0002
eval/loss : 1.5962
eval/f0 : 1.4615
eval/sil : 0.1347

--- epoch 100 ---
train/loss : 1.0587
train/f0 : 0.0843
train/sil : 0.9744
train/learning_rate: 0.0000
eval/loss : 1.6313
eval/f0 : 1.4785
eval/sil : 0.1528

Thanks~

[Charlottecuc]

The training loss looks quite different with the eval loss (especially the sil eval loss). Is it normal? 🤔

[yl4579]

I think this indicates overfitting. Either you add more training data, add data augmentation, or you should stop training before the evaluation loss starts to increase.

[Charlottecuc]

I think this indicates overfitting. Either you add more training data, add data augmentation, or you should stop training before the evaluation loss starts to increase.

How much data did you use to train the model (I used about 60 hours, 100 speakers)? Besides, why the eval f0 loss much higher (~1.4) than the training loss (~0.09)? Thank you very much.

[MMMMichaelzhang]

I use 128 speakers to train including singing and speaking.the sil eval loss doesnt seem normal. @yl4579 @Charlottecuc
Is it because I removed the silence part of all the audio?

[Charlottecuc]

I use 128 speakers to train including singing and speaking.the sil eval loss doesnt seem normal. @yl4579 @Charlottecuc Is it because I removed the silence part of all the audio?

Same loss curve. I used data augmentation (e.g. adding reverberation & noise) and the eval sil loss still starts increasing after epoch 30. Besides, when inferencing with sentences which have lots of plosive/fricative consonants, the predicted F0 values are not as good as that trained by @yl4579 .
Maybe something different in the trick of data augmentation?

[Ruinmou]

What type of data did you use for training? Can you provide me with a small sample? I used the speech data of more than 200 speakers to train the loss to a minimum of 3. After adding data enhancement (background noise), the loss dropped to 4.4 and it became very slow. @Charlottecuc @MMMMichaelzhang

[MMMMichaelzhang]

@Charlottecuc
Did you solve the sil loss problem?

@hai8023
I am using mandarin singing and speaking data，My data would make the sil loss go up and it doesn't look right.

[Charlottecuc]

@Charlottecuc Did you solve the sil loss problem?

@hai8023 I am using mandarin singing and speaking data，My data would make the sil loss go up and it doesn't look right.

I do not use the default f0 extractor (e.g. dio / harvest) offered by the author, and turn to other algorithms. Then the sil loss looks much better than before.

[skol101]

@Charlottecuc Did you solve the sil loss problem?
@hai8023 I am using mandarin singing and speaking data，My data would make the sil loss go up and it doesn't look right.

I do not use the default f0 extractor (e.g. dio / harvest) offered by the author, and turn to other algorithms. Then the sil loss looks much better than before.

@Charlottecuc did you still apply reverb/noises to the 30% of the dataset?

[Charlottecuc]

@Charlottecuc Did you solve the sil loss problem?
@hai8023 I am using mandarin singing and speaking data，My data would make the sil loss go up and it doesn't look right.

I do not use the default f0 extractor (e.g. dio / harvest) offered by the author, and turn to other algorithms. Then the sil loss looks much better than before.

@Charlottecuc did you still apply reverb/noises to the 30% of the dataset?

Yes. You can add some data aug after maybe 20 epochs.

[skol101]

@Charlottecuc Did you solve the sil loss problem?
@hai8023 I am using mandarin singing and speaking data，My data would make the sil loss go up and it doesn't look right.

I do not use the default f0 extractor (e.g. dio / harvest) offered by the author, and turn to other algorithms. Then the sil loss looks much better than before.

Did you use pyin/yin f0 extractor?