backbone

[wangmingaaaaa]

Hi, thank you for sharing the code. Does your train_diffusion_2d.py use unet as the backbone?

[McGregorWwww]

No, we use a 2D version of V-Net for train_diffusion_2d.py by directly changing the Conv3D to Conv2D.

[wangmingaaaaa]

Thank you very much for your reply. I want to compare your code as a baceline. I used unet as the backbone before, so I want to replace your backbone with unet. Do you think it is feasible? Or if I directly use your training results for comparison, is it unfair that the backbone is different?

[McGregorWwww]

Sure, it is very easy to change the backbone to unet, I can also give you a rough example in the attachment as reference (may exist bugs).

DiffUNet.txt

Furthermore, I've also tried unet before, but there's no big difference, so you may also use the results directly, I think it's fine.

[wangmingaaaaa]

Thank you very much, I will try the method you shared to change backbone to unet. I would like to ask, do you still have the results of using unet as the backbone to train on the M&MS data set? If so, can you share it?

[McGregorWwww]

Sorry, I did not save the results😅.

[wangmingaaaaa]

OK, thank you very much for your patient answer.

[ieeedre]

Hello, I conducted an experiment on the m&ms data set according to the DiffUNet.txt you provided. I can achieve an accuracy of about 72% on 2% of the D domain, and an accuracy of about 69% on the 2% of the A domain. , the effect is very different from what you said. I would like to ask if you set some specific parameters when using unet as backbone training?