DeweiHu / OCT_DDPM

Application of Diffusion Probablistic Model for unsupervised OCT denoising

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

如何运用在自己的图片数据集上进行运行

gguyinging opened this issue · comments

我不知道我的理解是否正确,我想我的训练集是原图,用来进行加噪扩散和逆向扩散,我的测试集则是有噪声的一些图片,是吗?我想我的训练集是4000张.png的图片,保存在一个路径中,测试集图片如上,我不知道该如何更改?以及您程序中E://中的tool我需要下载码?

Yes, you should add Gaussian noise with a beta schedule to the original images and train the denoise model. Note that in the testing phase you need to try different t values to find the optimal one. Please ignore the additional system path in my code (that was the directory of the helper function file util.py).

You can check out this #10

好的老师,就是在OCT——dataloader中践行更改吗?

老师,你好。我遇到了在自己的图片数据集上进行运行,但是其loss value在一百左右,想问一下正常应该是多少?过大或过小可能的原因是什么?

目前程序中的loss value是不是对plot出的图片进行一个验证的损失?我可以加大验证集的数量进行一个平均?

The plot is just a simple observation of the training process. The loss is computed by loss = noise_estimation_loss(model, x, t, e, b), the MSE between the Gaussian noise prediction model(x, t) and the real noise e.

我在想损失值几百,会不会是因为加载图片的时候没有除以255,所以过大?您的损失值大约在什么范围内?

For diffusion applications, the intensity is usually normalized to [-1, 1], otherwise the scheduled small Gaussian noise is not sufficient to convert the image distribution to the pure Gaussian distribution. In this work, I normalized to [1, 3].

所以对于我的图片数据集,应该将强度归一化为【-1,1】,我将做出尝试,谢谢您的建议

There is a more recent code for both unconditional and conditional diffusion here https://github.com/DeweiHu/Medical-Semantic-Diffusion/tree/main/src, I think it is cleaner. You can refer to it if needed.

好的,谢谢您