Question about the prepare_testset
mrgreen3325 opened this issue · comments
Hi, thanks for your work.
I find a problem when I run the prepare_testset.yml
There are several outputs for the same input with different sig1, sig2, theta setting.
May I know which one should be used to train my sr model or which can have the best quality?
Thanks.
As indicated by the name, prepare_testset.yml is used for generating testing set. For training, the HR-LR pairs are generated one-the-fly, which means only HR path is required.
As indicated by the name,
prepare_testset.ymlis used for generating testing set. For training, the HR-LR pairs are generated one-the-fly, which means only HR path is required.
Thanks.
Yes, I want to generate the HR-LR pair for my training program.
May I know which setting of this HR-LR pair should I use in my training?
We generate HR-LR pairs on-the-fly. The parameters on degradation are set at
We generate HR-LR pairs on-the-fly. The parameters on degradation are set at
Thanks for reply.
I follow the train_stage1.yml setting to config prepare_testset.yml to process 4x downscale as
name: 001_MANet_prepare_dataset
suffix: ~
model: blind
distortion: sr
scale: ~
gpu_ids: [6]
kernel_size: 21
code_length: 15
sig_min: 0.7
sig_max: 10.0
sig: 1.6
sig1: 6
sig2: 1
theta: 0
rate_iso: 0 # 1 for iso, 0 for aniso
sv_mode: ~
test_noise: False
noise: 15
datasets:
test1:
name: Set5
mode: GT
dataroot_GT: ../datasets/toy_dataset/HR_si
dataroot_LQ: ~
network_G:
which_model_G: MANet_s1
in_nc: 3
out_nc: ~
nf: ~
nb: ~
upscale: 0
#
path:
strict_load: true
pretrain_model_G: ../experiments/pretrained_models
However, the prepare_testset still produce so many different version of downscale LR.
Is that I miss something?
As shown in the Read.me. There are three settings.
1, for training, use train_stage*.yml. It will generate HR-LR pairs on the fly.
2, for testing, use prepare_testset.yml. It will generate and save different versions of LR (with different degradations) for testing.
3, for testing, you can also use test_stage3.yml. It will generate HR-LR testing pairs on the fly, but all of them follow the same degradation (different from case 2).
