yulunzhang/rnan

PyTorch code for our ICLR 2019 paper "Residual Non-local Attention Networks for Image Restoration"

rnan's Issues

Training Datasets

Thank you for the public implementation. Could you provide the links to the training data used for demosaicing, denoising and Compression Artifact Removal tasks.

the meaning of dir_demo,dir_test in option.py

Hi, excuse me. I have some questions that may need your help.

1. I know the dir_data is DIV2K 800 training images path. Is dir_demo set5,set14 test set path?
   Could you give a explanation of dir_demo ?

2. data_test: The data_test is default 'DIV2K'. And how many valid images used to test? Is it same as EDSR , i.e 0801.png to 0810.png?
   In other words , whether it lacks the parameter of data_range in option.py？

Best regards

Where is the Prepare_TestData_HR_LR.m?

Hello, I can't find Prepare_TestData_HR_LR.m.
Can you update it? Thank you!

Do we need `MSDataLoader`?

Hola friend, great code! is there a reason we need MSDataLoader instead of torch.utils.data.DataLoader?

image_SRF_2 or image_SRF_4?

Hicould you tell me the test_set Urban100 you used is image_SRF_2 or image_SRF_4?

How to deal with the case where the size of the image is odd

Hi
Since there is Subscale and Upscale in the mask module, how to handle the case where the size of the image is odd?

For pytorch>=1.0

Hi, thanks for your great work.
If it is possible, could you open the source code of pytorch >=1.0.0 that used more often.

Best regards.

hello，one small question about code

thanks for good work.

In this code，why “modules_body.append(conv(n_feats, n_feats, kernel_size))“ need to be add？if we remove it，the peformance will decrease？

some questions about parameters for training

Hi! Thanks for your wonderful work and open sources.
I'd like to reproduce the result of denoising. Here are some questions to consult:

1. What dataset is used for validation and how many iamges are used? (The readme mentioned that the validation set is div2k_val100, but in the options.py in the open source code, the default parameter shows that only the first 5 images of div2k_val_100 0801 ～ 0805 as used in the validation set)

2. What is the number of training epoches in the grayscale denoising task? How many days does the training take? (The parameter in options.py defaults to 1300 epoch, considering that the train set in each epoch is repeated 20 times, and the actual number of epochs is 26,000. We found that training 1/1300 epoch takes about 2 hours)

3. How to obtain the model corresponding the performance in paper? Did the test result shown in the paper use the best model on validation, or the model obtained from the last epoch?

4. Are the other parameters of the experiment to reproduce the result in paper the same as those given by options.py?

Looking forward to your apply. Thanks a lot!

Regards

The results of compression artifacts reduction are abnormal

Hi, thank you for sharing the code.

I run the code of CAR. The results are with abnormal intensity. Do you know the reason? Thanks

the difference between --data_test and --testset

Hi, thanks for your wonderful work.
If you don't mind, I hope you to explain the difference between --data_test and --testset.
When training and test the net, which one is used.
Thank you.

training time and number of GPUs

Hi, thanks for your wonderful work and opening source.
Could you please tell me how long did you train the model , the kind of GPU and number of GPUs?

Best regards

How much memory is required?

Thank you for your work. But when I try to run this code, it occur that "out of memory", I would like to know how much memory is required for this.

Thank you!

Can anyone reproduce the reported result?

Hi @yulunzhang ,

Thanks for your amazing work.
Could you please kindly provide the community more detail on how to reproduce your result?

We attempt to reproduce the experiment in gray image denoising with noise level equal to 50.
We use the same setting as you mention in the readme file and train the model three times.
However, the best result we can obtain is 26.35 instead of 26.48.

Another comment: the training time is around 8 days on one 1080Ti.

Very impressed work, here are the problems about experiments.

In Table 1, case 3 denotes that there exits NLB but without Mask Branch. According to my understanding, NLB could not be used alone. Could you guys give me some explanations? Thx :D

Problems about Kodak24.

Thanks for your impressive work! I want to know how did you get gray images of Kodak24? I trry opencv and matlab but can not get the same test results on Kodak24. Looking forward to your reply.

RuntimeError: CUDA out of memory.

I used the 1080Ti 11gb.
but when I run the t.test(),sr = self.model(lr, idx_scale)
I got the RuntimeError: CUDA out of memory.
how can I fix this? thanks

About the 'loss.step()' in Trainer.train(), trainer.py

# in code/trainer.py
def train(self):
        self.scheduler.step()
        self.loss.step()

Is the loss.step() only available for those loss functions that contain parameters update operating such as the loss of GAN?

Good perfomance but bad FPS

What is your inference speed? The FPS in my experiments with RNAN is really bad with default settings.