the-learning-and-vision-atelier-lava / dasr Goto Github PK

感谢您的工作，我有一个问题。当我用DF2K训练时，发现报错信息：

Traceback (most recent call last):
File "main.py", line 4, in
import data
File "/home/XXXX/F/XXX/DASR/data/init.py", line 2, in
from dataloader import MSDataLoader
File "/home/XXXX/F/XXX/DASR/dataloader.py", line 12, in
from torch.utils.data.dataloader import _DataLoaderIter
ImportError: cannot import name '_DataLoaderIter'

我的环境是：

torch 1.1.0
能解决下吗？

Hi. I'm sorry to bothering you.
Does this paper contain information about eigen values and rotate angles of the blur kernels in Tab.3 ?
I can see just shapes.

Thank you for sharing the exciting work. I have three questions:

1. In traditional blind-SR setting, only LR images are provided such as kernelGAN/ZSSR, while in your work HR images are used, and the network are trained using synthesized LR images in a supervised setting. I am wondering if it is blindSR actually.
2. Although multiple Gaussian blur/noise are added when generating LR images, it is still far from real-world degradation process. I am wondering if the model generalizes well to the real-world scenarios, where the degradation might not be Gaussian blur or noise.
3. Can your model work in real world unpaired LR/HR setting,where no systhesized images are used but unpaired LR and HR images are provided?

作者您好，我想问一下test时，也是先生成degradation representations再SR吗，还是直接对LR图片进行SR。

File "/home/XXX/anaconda3/envs/DASR/lib/python3.6/site-packages/numpy/lib/npyio.py", line 444, in load
raise ValueError("Cannot load file containing pickled data "
ValueError: Cannot load file containing pickled data when allow_pickle=False

感谢您在图像超分领域的贡献，您的工作对我有很大的帮助。请问下图6中用于提取degradation representations的moco是怎么训练的？是正常的训练还是只用了4种模糊核的数据进行训练？能提供下该部分的代码吗？

Hi, Thank you for your work and sharing code.

I would like ask to about multiscalesrdata.py code.
I don't understand why you crop HR images in line 157.
Can't we just crop the low resolution image for degradation learning?

Thank you!

Hello, thank you very much for your excellent work, but I would like to know how to train the model for realistic scene images without labels

如标题，在trainer.py里面没有看到有用到直接提供LR路径的地方，只有HR降质到LR再进行SR处理的过程，如果我想要直接输入LR图片进行输出该如何处理？
万分谢谢您的回复！

这是为什么啊

能否提供一下代码的环境,以及程序版本

Excuse me, the role of adversarial.py vgg.py in the entire network？ beginners hope you can help, thank you

您好 十分感谢您有理有据并且完整精彩的工作。不过有一点我还是不太理解，您提到的无监督是怎么体现的。或者说您认为他潜在的无监督价值怎么实现。可能是我论文读的不细致未能理解您的意图。十分感谢您的回答。

非常感谢您在图像复原领域做出的工作，对我的研究非常有帮助，然而我在用您在experiment中提供的model做测试时遇到了问题，由于使用的是个人电脑，用的是windows系统，不知道是不是在option中的参数设置有问题。

Traceback (most recent call last):
File "main.py", line 15, in
model = model.Model(args, checkpoint)
File "C:\Users\Luffy\Desktop\DASR-main\model_init_.py", line 29, in init
self.model = nn.DataParallel(self.model, range(args.n_GPUs))
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch\nn\parallel\data_parallel.py", line 142, in init
_check_balance(self.device_ids)
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch\nn\parallel\data_parallel.py", line 23, in _check_balance
dev_props = _get_devices_properties(device_ids)
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch_utils.py", line 455, in _get_devices_properties
return [_get_device_attr(lambda m: m.get_device_properties(i)) for i in device_ids]
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch_utils.py", line 455, in
return [_get_device_attr(lambda m: m.get_device_properties(i)) for i in device_ids]
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch_utils.py", line 438, in _get_device_attr
return get_member(torch.cuda)
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch_utils.py", line 455, in
return [get_device_attr(lambda m: m.get_device_properties(i)) for i in device_ids]
File "C:\Users\Luffy\Anaconda3\lib\site-packages\torch\cuda_init.py", line 312, in get_device_properties
raise AssertionError("Invalid device id")
AssertionError: Invalid device id
请问是GPU设置的问题吗？

感谢您的工作，请问可否提供一下，实验结果测试图集的原图（Figure 5，Figure 7，Figure 8），以及Figure 5中比较的两个核宽参数，以及4.4部分的真实图像的数据集。非常感谢，这会对我的研究有很大的帮助 麻烦您了！

作者您好，我想自己训练一个模型，对于训练退化表示的网络，我应该如何修改代码，为何k取32×256＝8192，其中256是什么？是32个退化的图片每个截取256个patch吗？感谢您的解答

Thanks for the great job. I had some trouble with my environment settings. but I could not find the environment settings requirement in your article or in your code ,could you please give me a brief introduction about you settings(GPU memory.)
Looking foward to your reply.Thanks again for your job.

根据您补充材料里面说的梯度和psnr的关系，我自己也选了几张图像进行了实验，并且用的模型是您代码里提供的只有同性高斯模糊核的模型，但是我测出来的结果，梯度和psnr的关系并不是很明显，所以想请问一下，您的平均梯度是怎么算的？是只有一幅图有这个结果还是在多幅上得出来的？在模型上有选择吗？想请问一下具体问下您怎么测的。

您好，代码可以在Windows下运行吗

作者你好，训练时每个epoch结束都会报错：NoneType object is not iterable，请问是我哪里设置错了吗？

Hello, thank you very much for your excellent work. Does the trainning data only need HR images? Can I delete the self.dir_lr in df2k.py ?

The resolution of images in DF2K is relatively high. May I ask if you need to crop to a certain resolution before training?

As mentioned in your paper, you "re-trained our DASR using their degradation model and
provide the results in the supplemental material" for comparisons with DAN/USRNet. Could you share these supplemental materials? Thanks

问题已解决

Thanks for the author providing interesting work!
Can you provide your checkpoint on 4x isotropic model? When we train your DASR as your setting on the paper, the performance of DASR is even lower than predictor+SRMDNF 0.8dB on Set5 under 4x iso! (We have train 5 parallel DASR models, and select the best one.) Now, my experiment shows that unsupervised representation has no superiority! Besides, the retraining predictor+SRMDNF is 1.7 dB higher than your paper shows on isotropic sigma=2.6 on Set14.
We wish the author to provide checkpoint on 4x isotropic model to help us find the issues.

您好，十分感谢您的工作带来的帮助。请问option.py文件中的test_every参数是在何处调用的，在训练时整个网络时如何实现每训练一个epoch就test一次呢？

您好，未找到您提供的模型，能否告知您提供的模型的位置，谢谢

您好！非常喜欢您的工作，我自己在两张1080ti上从头train的时候，发现显存占用只有759MB和691MB，但是程序也在正常输出训练过程，请问这正常吗？

Hi.

There's a problem when resume training.

I tried to restart training DASR using this :

python main.py --dir_data='my/path' \
               --model='blindsr' \
               --scale='4' \
               --blur_type='aniso_gaussian' \
                --noise=25.0 \
               --lambda_min=0.2 \
               --lambda_max=4.0 \
               --start_epoch=157\
               --resume=157\

The problem is that contrastive loss gets bigger.
I think parameters of encoder for degradation representation can't be loaded.

[Epoch 158]	Learning rate: 1.00e-4
Epoch: [0158][6400/31050]	Loss [SR loss: 9.753 | contrastive loss: 0.892 ]	Time [ 145.0 s]
Epoch: [0158][12800/31050]	Loss [SR loss: 9.747 | contrastive loss: 0.920 ]	Time [ 143.7 s]
Epoch: [0158][19200/31050]	Loss [SR loss: 9.722 | contrastive loss: 0.918 ]	Time [ 144.1 s]
[Epoch 158]	Learning rate: 1.00e-4
Epoch: [0158][6400/31050]	Loss [SR loss: 9.598 | contrastive loss: 7.457 ]	Time [ 145.2 s]

作者您好，请问您有做过关于对比学习中的K的设置，对模型性能有何影响的实验吗？

感谢您的工作。我在训练过程中，由于只有单GPU，所以在数据并行的部分出现了报错
通过查阅相关资料，这个问题仍没有解决，所以想向您咨询一下该如何处理这个问题

Hi. I'm sorry to bothering you.

Are the PSNR results calculated by MATLAB with YCbCr space or by your calc_psnr in utility.py?

Thanks!

Hi, Thank you for your work and sharing code.
Can the network model only input LR images for training?

Thanks for this interesting work!
Since only x2 isotropic model was released in "./experiment" folder, can you release the pretrained x3 and x4 isotropic models for testing?

Will you release pretrained models?

Hi, author. Thank you for your excellent work. But I'm a little confused about the code which calculate PSNR and SSIM metric. For PSNR, why the diff is multipide by a convert coefficient and shave the border when benchmarch is set to True. For SSIM, why not directly use api from skimage? Waiting for your response, sincerely.

Thank you for sharing your work! I have a question about the training steps:
When train the whole network, whether the parameters of the Degradation encoder network are fixed or not?

作者您好！我今天想继续训练，bash文件如下所示
python main.py --dir_data='./dataset'
--model='blindsr'
--scale='2'
--blur_type='iso_gaussian'
--noise=0.0
--sig_min=0.2
--sig_max=2.0
--save='bldsr_repro2'
--resume=319
但是输出文件显示从epoch1重新开始训练了。我找不到bug在哪里，请问这是为什么呢

您好 ，请问论文表2 中的kernel width 与test.sh 中 lambda_1,lambda_2的关系是什么呢？

我直接跑了您提供的代码，请问HR图和LR图需要自己提供吗，我只存放HR图，直接生成了*.pt文件 。

Hi. I'm sorry to bothering you.

How can i change the code for the latest torch version?

For torch==1.7.1 I think _DataLoaderIter matters.

Thanks for the impressive work!
To train the dataset, I downloaded 2650 Flickr2K HR images and 900 DIV2K_train_HR images, and copy them together to the /traing_path/HR folder. However, when I change the path and run the main.sh code, an error occurs in line 65 of multiscalesrdata.py: "self.repeat = args.test_every // (len(self.images_hr) // args.batch_size) ERROR: ZeroDivisionError: integer division or modulo by zero".
Is there anything wrong with the train set? Could you please offer some advice on the situation? Thanks a lot!

第一张图是我自己在X2 各向同性高斯核上的测试结果，配置信息如下： 为什么与您论文中的X2iso 测试结果有差距呢？
python main.py --dir_data='/home/bobo/F/SuperResolution/DASRend' \ --model='blindsr' \ --scale='2' \ --blur_type='iso_gaussian' \ --noise=0.0 \ --sig_min=0.2 \ --sig_max=2.0

Thank you for sharing your excellent work!!
In your test code, it seens we only need to provide benchmark HR test images, but how could I test on a real-world image that dont have label ?

感谢您的工作，我在训练中遇到一些问题，想向您请教一下
1.请问在选择不同的放大倍数，sig_min/max是否要设置不一样的范围？例如： [0.2,2.0], [0.2,3.0] and[0.2,4.0] for ×2/3/4 SR
2.在训练加 Noise的退化时，--noise需要设置成25还是其他值？
3.Table3中

对应的模糊核参数（致使其高矮胖瘦，旋转角度），lambda_1，lambda_2，theta，您分别设置的是多少？
在此表示十分感谢！

thank you for your good job. When i test your project ,i got an error

"ImportError: cannot import name '_DataLoaderIter' "

My enviroment is pytorch1.1.0 and python 3.6 .
what can i do?
thank you!