I added korean language code. How to train base model?

[jwj7140]

Hello. First, thank you for amazing project.
I added some code about supporting korean in this repository: https://github.com/jwj7140/Bert-VITS2-Korean.
After that, I've tryed to train model with train_ms.py. But it did not work well.

Using noise scaled MAS for VITS2
Using duration discriminator for VITS2
INFO:models:Loaded checkpoint 'Data/kss/models/DUR_0.pth' (iteration 0)
ERROR:models:enc_p.emb.weight is not in the checkpoint
ERROR:models:enc_p.tone_emb.weight is not in the checkpoint
ERROR:models:enc_p.language_emb.weight is not in the checkpoint
ERROR:models:enc_p.ko_bert_proj.weight is not in the checkpoint
ERROR:models:enc_p.ko_bert_proj.bias is not in the checkpoint
ERROR:models:emb_g.weight is not in the checkpoint
INFO:models:Loaded checkpoint 'Data/kss/models/G_0.pth' (iteration 0)
INFO:models:Loaded checkpoint 'Data/kss/models/D_0.pth' (iteration 0)

I think existing base model(like here) is useless in this case. So...I have to try to train new base model in korean.
Can you introduce the way of training base model in detailed?

[thundax-lyp]

this issue occurs when loading the model. see load_checkpoint() in utils.py.

assert saved_state_dict[k].shape == v.shape, (
    saved_state_dict[k].shape,
    v.shape,
)

maybe you are loading a wrong base model. you had better train the base model by yourself while add a new language

[jwj7140]

Yeah. I think so too. But despite many attempts to train new base model(with train_ms.py), it can't generate proper voice until now.
Here is output .wav file of my new base model in english: https://drive.google.com/file/d/1Pn0f3tuJJGUW_gSzVCg0QRZhks1o2mpt/view?usp=sharing

Are there some special ways to train base model? If not, how many epochs are required to train base model?