A Question on the Training Process
Hi, I trained my own model successfully and it worked out fine, but I have a question on the training process. I notice that the program (Colab code) actually starts to delete many different paths from one point (after it uploads some certain paths). I wonder what it is exactly doing. Do we need this process of deleting paths if we just want a sovits model? Thank you in advance!
By the word "paths", I mean the .pth files. Thanks.
Hey sorry for the late reply - missed this. What its doing is deleting the extra checkpoints so you don't end up with a massive amount of them in your repo. I did it because I felt it was wasteful to have that much in storage, but I guess it doesn't matter cuz its still there w/ lfs either way...
Got it, thanks for your reply! By the way, do you by any chance know the differences between sovits and Retrieval based Voice Conversion(RVC)? And I wonder if there is an easy way to train RVC like the notebook you created for sovits training. Thank you😄
I have no idea what the difference is 😅
I didn't try the RVC training since it looked a little complicated, but I'll have a try later. Also, I wonder what's the difference between G_10000.pth and D_10000.pth which are both the outputs of the sovits training. Thanks!
you only need the G-*.pth for inference. G stands for Generator, D for Discriminator. If you want to resume training, you'd need both.
Thank you!