Go to the rain training problem

[Qixin-oss]

Why do I train the rain removal model on 4090 and the psnr can only reach about 28?

[Algolzw]

We use PSNR_Y and SSIM_Y as the evaluation metrics. But the results should be higher (at least higher than 30 in terms of PSNR_Y) on the Rain100H dataset. Can you share your settings such that I can find the difference?

[Qixin-oss]

general settings

name: ir-sde
use_tb_logger: true
model: denoising
distortion: derain
gpu_ids: [0]

sde:
max_sigma: 10
T: 100
schedule: cosine # linear, cosine
eps: 0.005

degradation: # for some synthetic dataset that only have GTs

for denoising

sigma: 25
noise_type: G # Gaussian noise: G

for super-resolution

scale: 4

datasets

datasets:
train:
name: Train_Dataset
mode: LQGT
dataroot_GT: dataset/rain/trainH/GT
dataroot_LQ: dataset/rain/trainH/LQ

use_shuffle: true
n_workers: 4  # per GPU
batch_size: 4
GT_size: 128
LR_size: 128
use_flip: true
use_rot: true
color: RGB

val:
name: Val_Dataset
mode: LQGT
dataroot_GT: dataset/rain/testH/GT
dataroot_LQ: dataset/rain/testH/LQ

network structures

network_G:
which_model_G: ConditionalUNet
setting:
in_nc: 3
out_nc: 3
nf: 64
depth: 4

path

path:
pretrain_model_G: ~
strict_load: true
resume_state: ~

training settings: learning rate scheme, loss

train:
optimizer: Adam # Adam, AdamW, Lion
lr_G: !!float 1e-4
lr_scheme: MultiStepLR
beta1: 0.9
beta2: 0.99
niter: 700000
warmup_iter: -1 # no warm up
lr_steps: [200000, 400000, 600000]
lr_gamma: 0.5
eta_min: !!float 1e-7

criterion

is_weighted: False
loss_type: l1
weight: 1.0

manual_seed: 0
val_freq: !!float 5e3

logger

logger:
print_freq: 100
save_checkpoint_freq: !!float 5e3
In the settings of the IR-SDE