search

yl4579 / StarGANv2-VC

StarGANv2-VC: A Diverse, Unsupervised, Non-parallel Framework for Natural-Sounding Voice Conversion
MIT License
466 stars 110 forks source link

How much loss model will converge？ #5

Closed 980202006 closed 2 years ago

980202006 commented 2 years ago

hi, I want to train a new model using mixed open source data. I would like to know, How much loss model will converge And the required gpu resources and time-consuming

yl4579 commented 2 years ago

You may want to train with the provided data and config first and figure it out yourself because it varies between machines and configurations.

980202006 commented 2 years ago

If I just want to reproduce the training, use the data and model structure given in the example，How much loss model will converge？

yl4579 commented 2 years ago

As with training other GAN models, it is difficult to tell if the model has converged. You will need to inspect the generated samples regularly and pick the checkpoint that produces the best results. Unfortunately, loss values are not very interpretable in GAN training, but the cycle consistency loss can be used as an indicator for stopping training. If you see no significant decrease in cycle consistent loss in the validation set, it's probably time to stop.

The checkpoint we provided here was trained for 150 epochs, the same as in the provided configurations. However, it depends on the dataset and other settings, even though the models are quite robust against different datasets and parameters.

980202006 commented 2 years ago

Perfect answer！This is too helpful for practice.