text_ce and mel_ce loss

[weidezhang]

it seems when i train the model, the training loss_text_ce and loss_mel_ce_loss as well as validation loss val_loss_text_ce is trending down as expected but the other validation loss val_loss_mel_ce is trending up.

The training set and validation are quite similar. The data set is small though about 100 wavs each with length at most 8 seconds.

Any idea what might cause this ? Below is my experiment with val_text_ce loss and val_mel_ce loss.

Thanks,

[weidezhang]

gpt_gpt and gpt_encoding loss looks normal though

[152334H]

val_loss_mel_ce is probably the most important one. since you say training loss_mel_ce is going down, I would guess that your model is overfitting on the dataset

[FurkanGozukara]

what loss value is good? i written a script myself and prepared a very comprehensive tutorial
hopefully i will post soon on https://www.youtube.com/SECourses

so here my plot