想请问作者一些问题

[LiJiajun000322]

作者您好，我想请问在daclip-sde中，能否配置让daclip不工作呢，就完全是一个sde去工作，这可以在配置文件中实现吗

[LiJiajun000322]

general settings

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

sde:
max_sigma: 50
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: /home/image-restoration-sde-main/codes/datasets/low-light/trainH/GT/
dataroot_LQ: /home/image-restoration-sde-main/codes/datasets/low-light/trainH/LQ/

use_shuffle: true
n_workers: 20  # per GPU
batch_size: 16
GT_size: 256
LR_size: 256
use_flip: true
use_rot: true
color: RGB

val:
name: Val_Dataset
mode: LQGT
dataroot_GT: /home/image-restoration-sde-main/codes/datasets/low-light/testH/GT/
dataroot_LQ: /home/image-restoration-sde-main/codes/datasets/low-light/testH/LQ/

network structures

network_G:
which_model_G: ConditionalUNet
setting:
in_nc: 3
out_nc: 3
nf: 64
depth: 4
ch_mult: [ 1, 2, 4, 8 ]
context_dim: 512
use_degra_context: true
use_image_context: true

path

path:
pretrain_model_G: ~
strict_load: true
resume_state: ~

training settings: learning rate scheme, loss

train:
optimizer: AdamW # Adam, AdamW, Lion
lr_G: !!float 2e-4
lr_scheme: MultiStepLR
beta1: 0.9
beta2: 0.99
niter: 700000
warmup_iter: -1 # no warm up
lr_steps: [2000, 4000, 6000]
lr_gamma: 0.5
eta_min: !!float 1e-6

criterion

is_weighted: False
loss_type: l1
weight: 1.0

manual_seed: 3471
val_freq: !!float 1e4

logger

logger:
print_freq: 200
save_checkpoint_freq: !!float 1e4

我按如上配置文件去训练了一个lowlight的ir-sde，数据集就是LOL数据集。感觉效果不好，请作者指点一下配置参数

[LiJiajun000322]

还有如果是低光任务的话，应该选IR-SDE还是Denoising SDE呢，我注意到IRSDE的项目里有两种SDE定义

[LiJiajun000322]

Q3:daclip-sde文件做推理时，没有GT的情况应该怎么设置呢，我设置dataroot_GT:None会报错。
datasets:
test1:
name: Test
mode: LQGT
dataroot_GT: None
dataroot_LQ: datasets/universal/deg_type/LQ

是否时要修改mode:LQ？我这样试过也不行

[Algolzw]

低光任务的话就用ir-sde就好。不用daclip的话这两个参数：use_degra_context: true, use_image_context: true要设置为false。在推理时，你尝试下同时修改mode: LQ 和去掉dataroot_GT这一行呢（不行的话可以dataroot_GT也设置为LQ路径试试）。

[LiJiajun000322]

train:
optimizer: AdamW # Adam, AdamW, Lion
lr_G: !!float 2e-4
lr_scheme: MultiStepLR
beta1: 0.9
beta2: 0.99
niter: 700000
warmup_iter: -1 # no warm up
lr_steps: [2000, 4000, 6000]
lr_gamma: 0.5
eta_min: !!float 1e-6

logger:
print_freq: 200
save_checkpoint_freq: !!float 1e4

这部分的话迭代次数也需要这么多？我在LOL数据集上跑，发现打印的psnr几乎没变过，都是33

[LiJiajun000322]

不用daclip的话这两个参数：use_degra_context: true, use_image_context: true要设置为false。

我这样做了以后发现无法进行低光恢复

[LiJiajun000322]

use_degra_context: true
use_image_context: false
后可以进行低光恢复
如果use_degra_context设置为false，图像仍然是很暗的

[Algolzw]

我还以为要重新训练。用pretrained weights的话确实是必须设置这两个参数的。

[LiJiajun000322]

推理的话有没有办法不加载daclip模块呢，仅仅只有sde部分工作？

[LiJiajun000322]

还有就是我自己在LOL数据集上训练IR-SDE的时候，PSNR几乎与迭代次数无关
24-11-27 04:41:52.877 - INFO: <epoch: 82, iter: 10,000, psnr: 32.039530
24-11-27 05:51:09.814 - INFO: <epoch:165, iter: 20,000, psnr: 32.376172
24-11-27 07:02:31.526 - INFO: <epoch:247, iter: 30,000, psnr: 32.891167
24-11-27 08:11:48.979 - INFO: <epoch:330, iter: 40,000, psnr: 31.893832
24-11-27 09:21:05.904 - INFO: <epoch:413, iter: 50,000, psnr: 32.960719

[LiJiajun000322]

如果我这里GT设置为和LQ一样的话，测试结果就没法完成低光增强（ir-sde重新训练）

[LiJiajun000322]

general settings

name: ir-sde
use_tb_logger: true
model: denoising
distortion: lowlight
gpu_ids: [1]

sde:
max_sigma: 50
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: /home/datasets/low-light/trainH/GT/
dataroot_LQ: /home/datasets/low-light/trainH/LQ/

use_shuffle: true
n_workers: 20  # per GPU
batch_size: 4
GT_size: 256
LR_size: 256
use_flip: true
use_rot: true
color: RGB

val:
name: Val_Dataset
mode: LQGT
dataroot_GT: /home/datasets/low-light/testH/GT/
dataroot_LQ: /home/datasets/low-light/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: AdamW # Adam, AdamW, Lion
lr_G: !!float 2e-4
lr_scheme: MultiStepLR
beta1: 0.9
beta2: 0.99
niter: 700000
warmup_iter: -1 # no warm up
lr_steps: [2000, 4000, 6000]
lr_gamma: 0.5
eta_min: !!float 1e-6

criterion

is_weighted: False
loss_type: l1
weight: 1.0

manual_seed: 3471
val_freq: !!float 1e4

logger

logger:
print_freq: 200
save_checkpoint_freq: !!float 1e4

这是我的训练参数，可以麻烦作者看一下吗，IR-SDE用于LOWLIGHT任务

[LiJiajun000322]

name: ir-sde
suffix: ~ # add suffix to saved images
model: denoising
gpu_ids: [0]

sde:
max_sigma: 25
T: 100

degradation:
sigma: 25
noise_type: G # Gaussian noise: G

datasets:
test1:
name: Val_Dataset
mode: LQGT
dataroot_GT:/home/datasets/low-light/testH/LQ/
dataroot_LQ: /home/datasets/low-light/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: pretrained/ir-sde.pth

推理的时候就是用的GT和LQ都是LQ的路径

[LiJiajun000322]

在推理时，你尝试下同时修改mode: LQ 和去掉dataroot_GT这一行呢（不行的话可以dataroot_GT也设置为LQ路径试试）。

这样不行，会报缺少GT路径的错误

Traceback (most recent call last):
File "test.py", line 102, in
GT = test_data["GT"]
KeyError: 'GT'

[Algolzw]

ir-sde设置看着没什么问题呀，按理说训练测试也应该没问题。

[LiJiajun000322]

请问作者当时训练低光任务是用了哪些数据集呢，我看IR-SDE原始论文里是没做低光增强，在DACLIP里有用IR-SDE对LOL数据集的测试集做测试，可以提供一下数据集吗，我想训练一个用于低光增强的单个IR-SDE模型。

[LiJiajun000322]

或者说作者这里可否提供一下单个IR-SDE在低光增强上的权重文件呢，非常感谢您的帮助！

[Algolzw]

数据集可以在这里下载，IR-SDE的权重我没有保存，不过你重新训练就行，应该很快。

[LiJiajun000322]

我这边就是重新训练后觉得有问题，能否作者提供一下呢:D

[LiJiajun000322]

我发现我训练的模型，使用test给好LQ和GT的路径，模型能够正常输出恢复后的图像，但是用inference.py，或者把test的GT路径换成LQ，就没办法完成低光增强

[Algolzw]

按理说不应该啊。你直接把feed_data中GT设为None呢？或者检查一下图像保存的代码。