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

is something wrong with my StarGan loss curve? #57

Closed superhg closed 1 year ago

superhg commented 1 year ago

Hi, @yl4579 here is my training loss curve and eval loss curve, convert result is not good. The loss curve indicates overfitting?

image image

yl4579 commented 1 year ago

You may need to train for a longer time. 60 epochs are not enough for GAN to converge. Your cycle consistency loss is still decreasing.

skol101 commented 1 year ago

@superhg I wonder why you have eval/reg and eval/con_reg values? They are not used in the evaluation step as per code.