Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

why skip_diff_train before fast_epochs #12

Open
wblgers opened this issue Oct 22, 2023 · 7 comments
Open

why skip_diff_train before fast_epochs #12

wblgers opened this issue Oct 22, 2023 · 7 comments

Comments

@wblgers
Copy link

wblgers commented Oct 22, 2023

Hi,

Can you explain why skip diffusion train before the configured fast_epochs?
And how many epochs does diffusion train need?

Thanks!

@MaxMax2016
Copy link
Contributor

因为刚开始训练的时候,encoder还学的不好。encoder和diffusion都在拟合mel谱图,没有diffusion encoder也可以独立工作,整合系统甚至可以把diffusion完全删除也是可以的。等encoder学好了,在加入diffusion训练,可以节约训练时间。diffusion训两百个epoch可以看到效果。

@wblgers
Copy link
Author

wblgers commented Oct 24, 2023

因为刚开始训练的时候,encoder还学的不好。encoder和diffusion都在拟合mel谱图,没有diffusion encoder也可以独立工作,整合系统甚至可以把diffusion完全删除也是可以的。等encoder学好了,在加入diffusion训练,可以节约训练时间。diffusion训两百个epoch可以看到效果。

谢谢你的解答,另外有一个问题。为什么SpeakerClassifier中的说话人个数是写死的?line 292 of encoder.py
self.speaker = SpeakerClassifier( n_channels, 256, # n_spks: 256 )

@MaxMax2016
Copy link
Contributor

256是声纹向量的维度

@wblgers
Copy link
Author

wblgers commented Oct 24, 2023

256是声纹向量的维度

好的,谢谢解答!

@wblgers
Copy link
Author

wblgers commented Nov 3, 2023

@MaxMax2016 请教一个问题,你是多卡还是单卡训练的?
我自己写了一个多卡的代码,训练速度上来了,但是mel_loss收敛的并没有单卡好。
单卡训练到250epoch大概是:
Epoch 250: | spk loss = 0.337 | mel loss = 0.330 | prior loss = 1.139 | diffusion loss = 0.097
多卡训练到250epoch大概是:
Epoch 250: | spk loss = 0.335 | mel loss = 0.335 | prior loss = 1.135 | diffusion loss = 0.099

@MaxMax2016
Copy link
Contributor

MaxMax2016 commented Nov 3, 2023

单卡,改用多卡要调学习率的吧

@wblgers
Copy link
Author

wblgers commented Nov 3, 2023

单卡,改用多卡要调学习率的吧
调了学习率之后,直接跑飞了,melloss一直0.72不下降

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants