cyanzhao42 / metasolve

Meta-learning for Solving Inverse Problem

• meta_module.py, modules.py, utils.py, dataio.py, inner_modules.py are essential scripts for proposed method.
• Note: to run the scripts, one need to correct the path informatin in dataio.py with path to local CelebA dataset. (such correction may need to be done at many places, like denoiser/denoiser.py, etc.)

Inpainting task

• script_inpainting.py is sample experiment script for training for inpainting task
• eval_inpainting.py is sample evaluation script
• model files used are stored in model_zoo/inpainting_center, and model_zoo/inpainting_random
• UNet implementation can be found in UNet folder

Deconvolution task

• denoiser training code can be found in denoiser/, with trained denoiser stored in model_zoo/denoiser/dncnn.ckpt
• script_deblur_meta_fp.py is sample experiment script for training for inpainting task
• eval_deblur_ircnn_fp.py is sample evaluation script
• model files used are stored in model_zoo/deblur
• HQS related file in DPIR folder (both training and eval)
• red_fp.py is the evaluation script for RED (training of unroll RED can be done by slightly modify script_deblur_meta_fp.py)

Reference

• more reference can be found in the writeup
• a lots of reference are taken from: https://github.com/vsitzmann/siren , https://github.com/cszn/DPIR, https://github.com/vsitzmann/pytorch_prototyping