search

shaoyanpan / 2D-Medical-Denoising-Diffusion-Probabilistic-Model-

This is the repository for the paper "2D Medical Image Synthesis Using Transformer-based Denoising Diffusion Probabilistic Model".
https://iopscience.iop.org/article/10.1088/1361-6560/acca5c
MIT License
30 stars 9 forks source link

Model configuration for 3D Model #1

Open dmus opened 1 year ago

dmus commented 1 year ago

Nice work, will you also share the model configuration you used for the model in "Synthetic CT Generation from MRI using 3D Transformer-based Denoising Diffusion Model"?

shaoyanpan commented 1 year ago

Thank you!

That paper is still in the major revision of a journal so our lab hope to release the configuration and github after the full acceptance. However, if I am doing some 64x64x64 volume synthesis, I think I will try something like:

num_channels=64 attention_resolutions="32,16,8" channel_mult = (1, 2, 3, 4) num_heads=[4,4,8,16] window_size = [[4,4,4],[4,4,4],[4,4,4],[4,4,4]] num_res_blocks = [2,2,2,2] sample_kernel=([2,2,2],[2,2,2],[2,2,2],[2,2,2])

dmus commented 1 year ago

Thanks, and do you use mixed precision or not? It can make inference a lot faster, but I can also imagine high precision is needed for the diffusion models?