search

HJ-harry / DDS

Official PyTorch implementation for Decomposed Diffusion Sampling (DDS), presented in the paper "Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems" (https://arxiv.org/abs/2303.05754).
44 stars 3 forks source link

A small problem #1

Open Breeze-Zero opened 7 months ago

Breeze-Zero commented 7 months ago

https://github.com/HJ-harry/DDS/blob/b09673b146370c9de14be59be69929051e183fc8/solver_2d.py#L217

Thank you very much for your work. It is worth learning. I am testing it on my MRI data. Here AT(y) seems to be a complex number matrix, while x0_t is a real number matrix. Is this deliberate design?

Breeze-Zero commented 7 months ago

I later found a situation, in the multi-MRI reconstruction, your script is to input the full sample data merged by the coil for simulation, but in practice, we will first get the undersampled data and then do coil merging and input into the script. When I did this, I found that the reconstruction results are worse than the simulation results of the full sample data merged by the input coil. I suspect the AT A operation in bcg, which is not reversible without full sampled data. Since I am still a beginner, I can't analyze the specific reasons.