search

zoubohao / DenoisingDiffusionProbabilityModel-ddpm-

This may be the simplest implement of DDPM. You can directly run Main.py to train the UNet on CIFAR-10 dataset and see the amazing process of denoising.
MIT License
1.48k stars 156 forks source link

Bad results on medical images #24

Open mlszy928 opened 1 year ago

mlszy928 commented 1 year ago

I run the default code and achieved satisfactory results on Cifar10, but when I train the model on 2d CT images, the generated images are almost white or black by the model after 50 epoches. image

zoubohao commented 1 year ago

Hi

I think this issue is because your data is 2d CT images. In CIFAR-10 dataset, the channel of these images is 3. You can 1. convert the 2d CT images to 3 channel or 2. change the code to fit your task.

Notice: This code is suitable for learning the processing of DDPM but may not be a good repo to do some engineering tasks.

Good luck.

mlszy928 commented 1 year ago

Thank you for reply I have already converted the CT images to 3 channels... Maybe I should change the code for medical images.

yangfar commented 1 year ago

I run the default code and achieved satisfactory results on Cifar10, but when I train the model on 2d CT images, the generated images are almost white or black by the model after 50 epoches. image

Excuse me, have you found some DDPMs suitable for 1D or 2D data?

ming-transfer commented 8 months ago

Excuse me, I also want to know that DDPM is suitable for 1D signal ？is there some bro can tell me ？