learn_prox_ops
learn_prox_ops copied to clipboard
Unreasonable results of optimal_BM3D_experiment_a
Hi, Tim,
I find the optimal BM3D setting yields extremely poor results.
The code of
python src/experiment_deblurring.py with experiment_name=experiment_a image_name=barbara elemental.optimal_DNCNN_experiment_a
produces good results of
INFO - main - Input PSNR: 22.757000
INFO - main - Final PSNR: 26.032000
However, shifting DNCNN to BM3D prior by
python src/experiment_deblurring.py with experiment_name=experiment_a image_name=barbara elemental.optimal_BM3D_experiment_a
leads to extremely poor results
INFO - main - Input PSNR: 22.757000
INFO - main - Final PSNR: 5.494800
with verbose setting=1, I find the objective value explodes even in iteration 1
iter 1: ||r||_2 = 4199973199709416.500, eps_pri = 4199973199898.533, ||s||_2 = 27999821280051452.000, eps_dual = 61527351260.748, PSNR: 5.4948 dB
iter 2: ||r||_2 = 1238656270915858769903616.000, eps_pri = 1238656270938838138880.000, ||s||_2 = 8257708468100152428593152.000, eps_dual = 17468928722936238080.000, PSNR: 5.332 dB
I was wondering whether you can help me to figure it out? Thank you very much!