about grayscale range of data

[weilongMao]

Is there any requirement for the grayscale range of data?
Do I need to normalize the data to [0,1] ?
Thank you very much!

[DeweiHu]

Yes, the image intensity range should be normalized to [0,1]. Otherwise, the Gaussian noise specified in range [-1,1] will not be sufficient for diffusion. Dewei

…

________________________________ From: UJSdog ***@***.***> Sent: Tuesday, November 29, 2022 11:19 AM To: DeweiHu/OCT_DDPM ***@***.***> Cc: Subscribed ***@***.***> Subject: [DeweiHu/OCT_DDPM] about grayscale range of data (Issue #3) Is there any requirement for the grayscale range of data? Do I need to normalize the data to [0,1] ? Thank you very much! — Reply to this email directly, view it on GitHub<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FDeweiHu%2FOCT_DDPM%2Fissues%2F3&data=05%7C01%7Cdewei.hu%40vanderbilt.edu%7C5c52accb76b340ada85808dad22deb67%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C638053391918020558%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=8xJPb5Sx2qGIiVIj7nIrCetOZSUT1CkmW%2BsgygX36uI%3D&reserved=0>, or unsubscribe<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAL6WZQBP3NDFGYQXSOZDG2TWKY3LJANCNFSM6AAAAAASOXIELI&data=05%7C01%7Cdewei.hu%40vanderbilt.edu%7C5c52accb76b340ada85808dad22deb67%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C638053391918020558%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=99zMoqUyP%2BVjLbb%2Fc9d9lHMxHhzJlwlZIVoWNVUSHSs%3D&reserved=0>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[eat-orange1]

Hello, I see that your code normalizes the data to [-1,1], so how do you say normalize to [0,1]

[DeweiHu]

Hi, I updated the normalization code earlier this year, it will normalize the intensity to [-1,1] now.

[eat-orange1]

Thank you very much for your answer. I am also using normalization to [-1,1]. During the training process, I used a noisy OCT image with a size of 512 * 512, but the loss did not decrease or converge during the training process. Do you know the reason.

[DeweiHu]

Hi, I uploaded a checkpoint of my model in the folder ckpts. I normalized my instensity range to [1,3] when training this model. You can try if it works for you.

[Koid217]

非常感谢您的回答。我也使用规范化来 [-1，1]。在训练过程中，我使用了大小为 512 * 512 的噪声 OCT 图像，但在训练过程中损失没有减少或收敛。你知道原因吗。

Hello, I have such a problem now, do you know the reason？