search

KidsWithTokens / MedSegDiff

Medical Image Segmentation with Diffusion Model
MIT License
980 stars 147 forks source link

Poor performance using dpm-solver #134

Open RoboticsZhang opened 10 months ago

RoboticsZhang commented 10 months ago

After trainning for 17W steps, I can get good results without using dpm-solver in 1000 steps, as shown below: image However, when using dpm-solver in 20, 50, or 1000 steps, the results all look like the image below: image I wonder why dpm-solver generate so weird results. Looking forward to your reply, thank you!

ipwjincheng commented 4 months ago

hello,I met the same question as you,have you solved it?

RoboticsZhang commented 4 months ago

hello,I met the same question as you,have you solved it?

Yep, maybe there are some flaws in the implementation of DPM-solver in MedSegDiff. I used the official version and then solved it. It is worth noting that the official version takes the epsilon as the input, not the mean and covariance, which are the direct output of the neural network.

RoboticsZhang commented 4 months ago

hello,I met the same question as you,have you solved it?

I just found that the latest DPM-Solver-V3 have included a demo with openai/guided-diffusion. The link is https://github.com/thu-ml/DPM-Solver-v3 which could help you.

zxk72 commented 2 months ago

@RoboticsZhang Hello, how to use the official dpm-solver code? How did you modify dpm-solver to make the code run successfully? Is it convenient to publish your modified code? Or can I add you on WeChat? Looking forward to your reply, thank you very much!