Zero-DCE icon indicating copy to clipboard operation
Zero-DCE copied to clipboard

much brighter output image of my trained weights than the one produced by the released weights

Open huntkao opened this issue 3 years ago • 2 comments

Thanks for share the implementation.

When I use the model which was trained myself following the training step, my output images were quite bright (at ~30th epoch).

Do you have any idea for possible reason?

Appreciate.

huntkao avatar Dec 30 '20 09:12 huntkao

You may change the learning rate to 10-5 or change the value of exposure loss to 0.5 or 0.45. The results are related to the training data. So, if you change the training datasets, the configurations may need to be changed. Thanks.

huntkao [email protected] 于2020年12月30日周三 下午5:54写道:

Thanks for share the implementation.

When I use the model which was trained myself following the training step, my output images were quite bright (at ~30th epoch).

Do you have any idea for possible reason?

Appreciate.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/Zero-DCE/issues/17, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIEVQELAHUMUKSBTUQKEJVLSXL2GJANCNFSM4VOEXOTA .

Li-Chongyi avatar Dec 30 '20 10:12 Li-Chongyi

Hi, Li,

Thank you for quick reply.
I understand the 2nd solution.
But why and how does the learning rate affect the output brightness?

Thank you again.

Hunt

Chongyi Li [email protected] 於 2020年12月30日 週三 下午6:32寫道:

You may change the learning rate to 10-5 or change the value of exposure loss to 0.5 or 0.45. The results are related to the training data. So, if you change the training datasets, the configurations may need to be changed. Thanks.

huntkao [email protected] 于2020年12月30日周三 下午5:54写道:

Thanks for share the implementation.

When I use the model which was trained myself following the training step, my output images were quite bright (at ~30th epoch).

Do you have any idea for possible reason?

Appreciate.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/Zero-DCE/issues/17, or unsubscribe < https://github.com/notifications/unsubscribe-auth/AIEVQELAHUMUKSBTUQKEJVLSXL2GJANCNFSM4VOEXOTA

.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/Zero-DCE/issues/17#issuecomment-752409064, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAJTDPMOKYJLQXSPSOHDI7DSXL6THANCNFSM4VOEXOTA .

-- -- hunt

huntkao avatar Dec 30 '20 10:12 huntkao