Closed a897456 closed 1 month ago
Hi, I don't preserve the results. However, your results seem reasonable. The mel loss will become higher in the GAN training stage.
However, your results seem reasonable.
Both are normal. GAN training is not stable since we alternatively update the generator and discriminator.
- I think there is a second problem with this graph, that is, neither the loss curve of the generator nor the loss curve of the discriminator has tended to be constant. Should the step size be increased and the training continued until the curves tend to be stable?
Do you think it is necessary to continue increasing the step size for training in this issue?
You can try it. I didn't find significant perceptual differences for training longer.
Hi @bigpon Can you show the training process graph of 'symAD_libritts_24000_hop300' ? It is implemented through the tensorboard command. Because I found that my training graph is chaotic, as shown in the following figure.![image](https://github.com/user-attachments/assets/75fda0a3-3e44-43b2-96a0-efcdf5bf26e7)