how many epochs get the good trained model?

[alexliyang]

I use FiveK data set , resize the jpg to 480P , and use expertC as the target, but after 20 epochs , the test result is not good .

how many epochs you trained while you get the good model?

[creotiv]

Can you she image output that you have?

…

On Fri, Jan 21, 2022, 3:14 AM alexliyang ***@***.***> wrote: I use FiveK data set , resize the jpg to 480P , and use expertC as the target, but after 20 epochs , the test result is not good . how many epochs you trained while you get the good model? — Reply to this email directly, view it on GitHub <#22>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAB5CDOOV6B6FVUPLZNFKFTUXCXOHANCNFSM5MOHZIXA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[alexliyang]

the network blocked upload image :(

the train data I used likes trainning 3DLUT model, 480p jpg file as the original , and use expertC's as the ground truth, then use this repo code train the hdrnet model.

is this dealing right?
I will try upload later

[alexliyang]

here is the goundtruth:

and the inference out is

I think they have more difference, how to improve the training model?

Can you she image output that you have?
…
On Fri, Jan 21, 2022, 3:14 AM alexliyang @.> wrote: I use FiveK data set , resize the jpg to 480P , and use expertC as the target, but after 20 epochs , the test result is not good . how many epochs you trained while you get the good model? — Reply to this email directly, view it on GitHub <#22>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAB5CDOOV6B6FVUPLZNFKFTUXCXOHANCNFSM5MOHZIXA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub. You are receiving this because you are subscribed to this thread.Message ID: @.>

[creotiv]

The problem here that you never get same result as ground truth, because of variance in editing even with the same author.

I also have problems with this implementation, cause training somehow different from original. Many people tried to understand why, we first though the problem in grid_sample function, but then found that it work correctly. so really here are still many questions

[creotiv]

I have smaller test dataset for such models, you can try to use it https://www.kaggle.com/anikishaev/photo-retouch-sm-ds
But variance there much bigger

[creotiv]

ive added bilateral_slice from original repo compiled for jit. But still has some problems with optimization for some reason.
So i think grid_sample was working correctly

[Varato]

I just noticed another thing from the original paper.

Where the paper states the PSNR is 28.8, they are talking about the HDR+ task. In the HDR+ task, they did NOT use the expert-retouched photos as the ground truth. Instead, they use the HDR+ ("a complex hand-engineered photographic pipeline that includes color correction, auto-exposure, dehazing, and tone-mapping.") to process the raw photos in the FiveK to get the ground truth to train the model. On the other hand, for the "Learning from human annotations" task, they did not mention PSNR any more, but the L*a"b color error alone.

So my guess is that the HDR+ pipeline outputs much more consistently styled photos than human retouching. So maybe we shouldn't expect training on the FiveK with expert_* as GT can hit the high PSNR.

Trying to verify this theory still (by utilizing some ISP pipeline to get consistent GTs).

[creotiv]

ive tested both networks original and mine on same datasets. Original working, mine for some reason not. You dont need HDR+ for the task, it works pretty good on simple photos with editors. Kindly yours, Andrey Nikishaev Areas ML/DS/CV/Soft Dev/BizDev/Growth Hacking/Customer Rel/IT LinkedIn http://ua.linkedin.com/in/creotiv GitHub http://github.com/creotiv Slideshare https://www.slideshare.net/anikishaev/ AnimalHelp https://patreon.com/uah Telegram @anikishaev Mobile +380632410666

…

On Thu, May 26, 2022 at 8:56 AM Xin Chen ***@***.***> wrote: Hi, thanks for reply. I just noticed another thing from the original paper. Where the paper states the PSNR is 28.8, they are talking about the HDR+ task. In the HDR+ task, they did NOT use the expert-retouched photos as the ground truth. Instead, they use the HDR+ ("a complex hand-engineered photographic pipeline that includes color correction, auto-exposure, dehazing, and tone-mapping.") to process the raw photos in the FiveK to get the ground truth to train the model. On the other hand, for the "Learning from human annotations" task, they did not mention PSNR any more, but the L*a"b color error alone. So my guess is that the HDR+ pipeline outputs much more consistently styled photos than human retouching. So maybe we shouldn't expect training on the FiveK with expert_* as GT can hit the high PSNR. Trying to verify this theory still (by utilizing some ISP pipeline to get consistent GTs). — Reply to this email directly, view it on GitHub <#22 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAB5CDPGTJUSTI7BHT2HNVTVL4HAHANCNFSM5MOHZIXA> . You are receiving this because you commented.Message ID: ***@***.***>