yulunzhang / rdn Goto Github PK

hi @yulunzhang
thanks for your great work
any chance to see tensorflow implementation for this project any time.

I 'am test on 6000x4000 pixel on Pre-train model out of memory. But EDSR no have problem.

->Torch7 on windows Nvidia 1080ti

您好。打扰了。我记得您之前提供过中文注释版本的code，请问可以再上传一次嘛？多谢多谢！！

The pytorch version of the code reproduces results that are different from those in your paper, and lua is rarely used nowadays. Can you please provide the super-resolved results of each dataset?

Hi
Thank you very mutch for your code. I have review the model. While, I am not familiar with torch and I get the result of 36.2648. When I read all the issues, I had found that you have skipped some batches when training. Could you tell me where the code or the regular. Waiting for you request.Thank you!

hi，sorry to bother you,but I have a problem that model does not converge,what should I do?can you give me some suggestions?thank you so much.the loss function i used is l1 loss

Hi, your work was excellent and I'm tring to reimplement it as I can understand it deeperly.

I have one question that you said in your paper "Learning rate decreases half for every 200 epochs", and you trian RDN "takes 1 day with a Titan Xp GPU for 200 epochs", did that mean you didn't half down learning rate in whole training? Cause as far as I understand you just train your net for 200 epochs.

Btw, do you think DATA AUGMENTATION is necessary for SR tasks? if input always a patch, 800 images of DIV2K can produce enough patches for 200 iterations of 1000 steps per iteration

Looking forward your reply!

Hi, Really loved reading the paper!

I just had one small confusion. What is MemNet lacking that you are trying to resolve?

Your point from the paper:
MemNet interpolates the original LR image to the desired size to form the input. This preprocessing step not only increases computation complexity quadratically but also loses some details of the original LR image.

I have read MenNet they also do have the same feature extraction initial network as yours followed by memblocks and reconnet.

So what does this above point mean?

Thank you!

Hi, thanks for you great work. I find nEpochs is set as 10000 in the opts.lua, does it mean that we need 10000 epochs to get the results in the paper ?

您好，本人没有学过lua语言，尽力去看懂您的代码，但还是不能理解。最初densenet在33之前加了一层11（减少输入特征图的数量，从而提高计算效率），看您论文当中并没有类似的操作，不知道您的代码或者实现中，是不是没有Bottleneck layers设计？非常抱歉用中文向您提问，感谢

Hi, Thank you very much for sharing.
I hava some questions in training.
How much time do you spending?
How many epoch you have been trained?

Hi, yulun,
Thanks for your excellent work.

I am wondering how to build the model you use? There are many files in "./models", like "RDN", "baseline". If I would like to construct the basic network you use as "Residual Dense Network", how should I follow each steps?

Thank you. :)

Do you have any results on DND (Darmstadt Noise Dataset), say sRGB ?

Hi, is image restoration already included in this repository? If so, it will be very kind of you if you could offer some guides on training and testing a model by myself. Thanks!

Hello, I have a problem. When I use RDN as the generator for training, the details of the generated image will appear R, G or B color spots. Have you encountered similar problems. BI down sampling is used for low resolution images, and this problem will not occur using residual network.

请问PTAMI模型的参数量是多少？您的CVPR论文的参数量好像与TPAMI的论文不同。PAMI中您好像只提及了x2的是22M，我想问一下x3和x4的

Hi, thank you so much for the source code. Besides, could you please provide the resulting image of the degradation model? I am not familiar with torch. I test the tensorflow version and find it do not reach the best performance.

Hi,
How much loss(MSE) did u obtained after your training got completed, I've trained my model for about 100 epochs and its ~100 ?
Moreover, I'm using RandomCrop of patches of LR and HR for training ...how will it effect my training?

Thank you

Hi,
I cannot install Torch7 because it is too old to install corresponding software package.
So do you have RDN models on the pytorch framework?
Or could you share some SR results produced by RDN?
I just want to compare visual quality.
Looking forward to your reply

linux or windows?

When I am working on '.lua' code to convert from png to t7 format it is showing error. the error is

Error -> [string "png_to_t7.lua"]:1: module 'image' not found:
no field package.preload['image']
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm\lua\image.lua'
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm\lua\image\init.lua'
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm\image.lua'
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm\image\init.lua'
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm..\share\lua\5.3\image.lua'
no file 'C:\Program Files\WindowsApps\61954lingguang.LuaScriptTest_2.0.0.0_x64__5b5h4n7n89bzm..\share\lua\5.3\image\init.lua'
no file '.\image.lua'
no file '.\image\init.lua'

I changed the path of DIV2K folder. still same error appears can any one suggest me what mistake i made.#

the code file attached below for your reference.

Thank you in advance.

Hello. I have a question.

I want to train the network by using SR291 dataset.
I tried to change the parameters and code.
However, when I train with 1 epoch, the validation result is too good (e.g. 37.0581).
It seems to be used some other model which I didn't train.
So, when I want to make new model with other dataset, where should I change?

Hi,

Thanks for your amazing work,

After checking the repo of EDST - pytorch, it seems that they didn't put denoising model in their code.

Therefore, I was wondering if anyone could provide it in PyTorch version.

Thanks in advance.

Hello,

I'm trying to replicate your model. did you remove all bias from your model ?

Hi,
I am interested in the NTIRE2018 track4 challenge (wild SR)
This may be a basic question but to which of your degradation methods does this challenge refer? (BI\BD\etc.)
To the best of your knowledge\assumption - would your network perform good SR on any random image (e.g. taken from a phone) or are there degradation limitations?

Thank you

th main.lua -scale 3 -netType resnet_cu -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 96 -dataset div2k -datatype t7 -DownKernel BI -splitBatch 4 -trainOnly true
我跑这个代码，但是160000次迭代也没有停下来，该怎么设置？我们需要多少迭代次数

Hello~
The test data can't download from BaiDu, could you share another link?
Thank you!