Questions about results mismatch with the numbers in the paper

[bznick98]

Hi authors, thanks for this innovative work! I tried to re-train the InitModel from scratch, and use that to re-train the Finetune Model from scratch as indicated in the paper. I used the exact same configurations (configs/init_neurop.yaml and configs/train/train_neurop_mit5k_dark.yaml) and dataset provided in the github repo, but the results seems a little bit off from what paper states. Would you mind me asking if there is any extra steps when you train the InitModel and FinetuneModel?

The results I get for MIT5K Dark Dataset is: Average PSNR: 23.904039526530063 SSIM: 0.8980636596679688 deltaE: 10.479456901550293

[amberwangyili]

our initialization scheme used total 500 images in dataset-init （we only provide half of the dataset since some of them are impaired while migrating to a new cluster）. If you want to get the test result，feel free to check the pretrained model provided in the repo. Zongnan Bao ***@***.***>于2023年5月12日 周五02:57写道：

…

Hi authors, thanks for this innovative work! I tried to re-train the InitModel from scratch, and use that to re-train the Finetune Model from scratch as indicated in the paper. I used the exact same configurations ( configs/init_neurop.yaml and configs/train/train_neurop_mit5k_dark.yaml) and dataset provided in the github repo, but the results seems a little bit off from what paper states. Would you mind me asking if there is any extra steps when you train the InitModel and FinetuneModel? The results I get for MIT5K Dark Dataset is: Average PSNR: 23.904039526530063 SSIM: 0.8980636596679688 deltaE: 10.479456901550293 — Reply to this email directly, view it on GitHub <#4>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AH3SV5RMPE5LMXPGPGMVP73XFUZCFANCNFSM6AAAAAAX6RDZCI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[bznick98]

Hi, is it possible to provide the entire 500-image dataset for InitModel? I really appreciate the time and the effort, thanks!

[cxzhou95]

our initialization scheme used total 500 images in dataset-init （we only provide half of the dataset since some of them are impaired while migrating to a new cluster）. If you want to get the test result，feel free to check the pretrained model provided in the repo. Zongnan Bao @.>于2023年5月12日 周五02:57写道：
…
Hi authors, thanks for this innovative work! I tried to re-train the InitModel from scratch, and use that to re-train the Finetune Model from scratch as indicated in the paper. I used the exact same configurations ( configs/init_neurop.yaml and configs/train/train_neurop_mit5k_dark.yaml) and dataset provided in the github repo, but the results seems a little bit off from what paper states. Would you mind me asking if there is any extra steps when you train the InitModel and FinetuneModel? The results I get for MIT5K Dark Dataset is: Average PSNR: 23.904039526530063 SSIM: 0.8980636596679688 deltaE: 10.479456901550293 — Reply to this email directly, view it on GitHub <#4>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH3SV5RMPE5LMXPGPGMVP73XFUZCFANCNFSM6AAAAAAX6RDZCI . You are receiving this because you are subscribed to this thread.Message ID: @.>

If so, please release the pretrained init model file, so we can train the network based on it to reproduce your paper results.