sczhou / lednet Goto Github PK

I want to export the model using onnx,I met an error when export the model,as onnx cann't support dynamic conv.Do you have any suggestions about solving this error?

There is no file named version.

The path in argument parse, file generate_low_ligh_imgs.py:
parser.add_argument('--test_path', type=str, default='/mnt/lustre/sczhou/datasets/deblurring/LOLBlur/high_sharp_scaled') should be generic and not include the developer's filesystem path.

If those arguments are required, it's a good practice not to include a default value, or provide a value that will work in all envirnoments.

Hello, thanks for your works.

I wanted to train the Conditional Zero-DCE on a new dataset but finally failed after several attempts. The structure of Conditional Zero-DCE has some inconsistencies with the original model, and I don't know how to make the loss functions work.

Could you please release the training code of Conditional Zero-DCE? I'm looking forward to your reply.

According to your paper, the location of the red dot is the saturated region, not the unsaturated region, right?

Sorry I linked the red dot to the red curve, the picture is not wrong.

Hello,

I'm trying to get the metrics shown in the paper (NRQM NIQE) but could not.
I used pyiqa library and tested with 3 different weight files provided in this repo.

Would you happen to know why? Thank you!

This is a great work!
I'm very interested in the data of LOL-Blur without synthetic blur and low light processing and wondering if you could share it.

I get 404 while downloading lednetgan.pth using included python script. and on Releases section this model does not exist. Where can i download this model?

Hello, your network, lednet.pth and lednet_ retrain_ 500000. pth These two networks are both through train_ Is LEDNet.yml the network for training?

What is the difference between the two? What adjustments are needed to obtain the two network models?

What type of machine was used in your training model at that time, and how many days would it take to train？
thank you

Different from Zero-DCE that uses different curves for RGB channels, CurveNLU applies the same curve to different channels in the feature domain.
Are you fitting all the pixel points with a curve to get the best curve parameters?
I don't really understand that CurveNLU applies the same curve to different channels in the feature domain.
Could you explain this knowledge?
Sorry for any interruptions.
thank you

could you please release the code of data synthesis pipeline?you have release the code of Darken,but seem to no other part of the pipe

Great work!
I am very interested about the raw data which are untreated by ISP. Could you release and upload them?

Hi, do you provide the blur kernel for this dataset? I found some blur parameters in the JSON file but I am unsure how they are used.

Thank you!

I followed the step-by-step installation, and did everything needed including installing pytorch and downloaded the pre-trained model.

As I'm trying the very first test:

# test LEDNet (paper model)
python inference_lednet.py --model lednet --test_path ./inputs

I get this error:

python inference_lednet.py --model lednet --test_path ./inputs
Traceback (most recent call last):
  File "inference_lednet.py", line 54, in <module>
    ckpt_path = load_file_from_url(url=pretrain_model_url['args.model'],
KeyError: 'args.model'

Maybe I missed something in the installation?
Can you please explain how do I fix that?
Thanks ahead! 🙏

wonderful work，please release dataset and code soon!!...

There is a 404 (not found) error when using download_pretrained_models.py.
Command: python3 download_pretrained_models.py LEDNet

Result: