guoqincode / Open-AnimateAnyone

Unofficial Implementation of Animate Anyone
2.9k stars 233 forks source link

about loss #24

Closed 21-10-4 closed 9 months ago

21-10-4 commented 9 months ago

image Why my loss is quite strange today, I try the new code, and my loss gets NaN: image

zanghyu commented 9 months ago

I have the same issue. I'm using Tiktok dataset.

21-10-4 commented 9 months ago

I also use Tiktok dataset

Bismuth209 commented 9 months ago

What's your learning rate?

kecol commented 9 months ago

Try decreasing the learning rate to something like 5.e-6

zanghyu commented 9 months ago

Try decreasing the learning rate to something like 5.e-6

I have tried 1e-4, 1e-5, and 5.e-6. None of them work... Does anyone have any idea?

guoqincode commented 9 months ago

Are you using the latest training code?

tdlhyj commented 9 months ago

Have you checked if the submitted code can train correctly? I've been using the latest code, but the loss consistently fails to decrease. Is it possible that some code has not been submitted? @guoqincode

guoqincode commented 9 months ago

Have you checked if the submitted code can train correctly? I've been using the latest code, but the loss consistently fails to decrease. Is it possible that some code has not been submitted? @guoqincode

I had no problem training on 8*A100.

Bismuth209 commented 9 months ago

Can you share the loss curve and some sample results?

On Mon, 18 Dec 2023, 12:37 qguopku, @.***> wrote:

Have you checked if the submitted code can train correctly? I've been using the latest code, but the loss consistently fails to decrease. Is it possible that some code has not been submitted? @guoqincode https://github.com/guoqincode

I had no problem training on 8*A100.

— Reply to this email directly, view it on GitHub https://github.com/guoqincode/AnimateAnyone-unofficial/issues/24#issuecomment-1859666477, or unsubscribe https://github.com/notifications/unsubscribe-auth/BEVOQCZS7BZ55XLPV6Z3FITYJ7TTVAVCNFSM6AAAAABAWB3ZBSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNJZGY3DMNBXG4 . You are receiving this because you commented.Message ID: @.***>

guoqincode commented 9 months ago

Can you share the loss curve and some sample results? On Mon, 18 Dec 2023, 12:37 qguopku, @.> wrote: Have you checked if the submitted code can train correctly? I've been using the latest code, but the loss consistently fails to decrease. Is it possible that some code has not been submitted? @guoqincode https://github.com/guoqincode I had no problem training on 8A100. — Reply to this email directly, view it on GitHub <#24 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/BEVOQCZS7BZ55XLPV6Z3FITYJ7TTVAVCNFSM6AAAAABAWB3ZBSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNJZGY3DMNBXG4 . You are receiving this because you commented.Message ID: **@.***> My current machine does not have access to external networks. I will organize the current repo after all the models are trained.