请问为什么我直接加载您在文件夹checkpoint下的模型进行测试，结果只有8db，是因为你放的是仅初始化的模型吗？

Hi Zheng,

This a great repo. Appreciate for letting us use it. I have 2 questions regarding training and testing.

1-) Why do we use testing data (testing folders) as validation data during training ?
In train_IMDN.py:

2-) Can we test our LR images without corresponding HR images? Ideally, we want to generate HR images from LR images? Do we have to have --test_hr_folder in order to test LR images?

python test_IMDN.py --test_hr_folder Test_Datasets/Set5/ --test_lr_folder Test_Datasets/Set5_LR/x2/ --output_folder results/Set5/x2 --checkpoint checkpoints/IMDN_x2.pth --upscale_factor 2

Hi, thank you so much for the source code. Besides, could you please provide the Memnet resulting image of the degradation model?

Hello! I want to ask how you caculate the parameters of IDN? I use torchsummaryX to caculate the model size, but the parameters is about 677k.

Thanks for release amazing work
Is ‘Training dataset DIV2K’ same with 'Train Data (HR images)' in https://data.vision.ee.ethz.ch/cvl/DIV2K/??
Can you provide the baidunetdisk of ‘Training dataset DIV2K’，thanks

您好，请问下 您是怎们把pth 转换成 tflite 格式的的，有代码或者教程链接么，谢谢

Hi,Zheng!
Could you give a brief introduction about how to change the pytroch model to the tflite model?
Should I use this routine "pytorch->onnx->tensorflow->tflite"?

Thanks for your attention!

where is the paper? Thanks

Dear Zheng：
您好！
我在运行test_IMDN.py时，遇到两个问题：
1）我可以运行成功您提供的Set5文件夹下.bmp数据集，然而我自己再测试Set14时，无论是.bmp还是.png格式下的数据集，都会运行出错。 【TypeError: ‘NoneType’ object is not subscriptable】 自己调试程序，好像是并没有读取相应的图片，麻烦您能解答一下是为什么吗？非常感谢
2）我按照您checkpoint文件夹下的.pth文件（X2\X3\X4）去测试test_IMDN.py ,然而得出的PSNR和SSIM都低于您论文里的值，我也用生成的SR图片和原有的HR图片在matlab里面计算PSNR和SSIM，也都低于您论文里的值，并且我也从头到尾跑过了train_IMDN.py，也生成了相应的.pth文件。

  非常接待您的回复和解惑，谢谢！

Hi,Zheng!
I trained the IMDN in a regular manner with the lr 2e-4.The lr reduced by half per 200 epochs for 1000 epochs totally.However,when I tested the network ,the results i got were lower than yours.After that, I increased the epochs to 300 when the lr is 2e-4 and 1e-4 and the total epochs was increased in 1200. But finally, the results were still lower than yours.The psnr and ssim i got were obtained by matlab in y channel.The experiment results i tested are as followed :
para|MAds| Set5 | Set14 | B100 | Urban100 | manga109 |
715 | 40.9 | 32.14/0.8942 | 28.53/0.7806 | 27.53/0.7351 | 26.00/0.7828 | 30.37/0.9067
So, could you tell me what i should do to further improve the final results?

OSError: Failed to interpret file '/media/npu/Data/mdx/DIV2K_train_LR_bicubic/X4/0710x4.png' as a pickle
That's why?

你好，首先非常感谢您的开源工作，在阅读文章的时候对于其中的adaptive cropping strategy的内容理解的还不够透彻，想请问一下，ACS的具体作用是什么呢，使用ACS之后为什么就可以处理arbitrary size和any scale factor 了呢，如果有时间希望您可以帮忙解惑一下，感谢

Thank you for nice work and thank you for sharing the code.

python train_IMDN.py --root /path/to/DIV2K_decoded/ --scale 4 --pretrained checkpoints/IMDN_x4.pth

Why do we need the pretrained model for 4x when we train the x4 SR model?

How to train a model that has similar performance with the paper? Thanks.

hello, i have a issue about test time.
i test the code with the Set5 by i5 cpu, but i get the time is about 4.4s;
and i know the cpu time will big but i think this time is too big, right?
because the flops is smaller than fsrcnn and almost models, and the gpu time is the same as others, even though the depth is different.

the reason is the depth?? if really it is, why the gpu time is still small?

thanks for the wonderful concept.
can u please let me know exactly where in the code is this strategy of adaptive cropping

hello, Zheng. I want to ask you why the last row of the code in function CCA is "return x * y " ,isn't "x + y "?

I want to know how to measure the memory consumption of model in your paper.
Thank you!

In your paper, the figure 9 shows that the speed of IDN is slower than the CARN and IMDN on Set5 for x4 SR.
However, I use the official code such as IDN, CARN and IMDN to evaluate the average inference time on Set5 x4 dataset. The average running time of these methods are 0.007s, 0.028s and 0.029s, respectively.
I find that the speed of IDN is faster than CARN and IMDN, and the running time of IMDN is close to CARN. I an confused about the result. Could you tell me the reasons?
My operation environment is as follows:

GPU: GTX 1080Ti
OS: Ubuntu 18.02 LTS
CUDA: 10.0
CUDNN: 7.4
Python version: 3.6
pytorch version: 1.0

Thank you！

wheather is the psnr in these pictures based on y-channel or RGB-channel?

hello,nice to meet you
I want to ask that
is checkpoint a pre-train model?
and how can i train it? thanks
looking forward your reply

I am interesting wheter video super resolution like EDVR could be applied using your distillation work?

Thx,
Lei

I tested IMDN_X4 using the provided checkpoint and Set5 dataset. The PSNR is 32.19. I also trained the IMDN_X4 from scratch strictly following your code. The best PSNR is 32.14. These two results are slightly lower than the reported 32.21 in Table 5.

What's the possible reason for the two differences? Do I need to train a model based on your pre-trained checkpoint? (i.e., --pretrained checkpoints/IMDN_x4.pth) If so, how was this checkpoint trained?

Is there any plan for the release of the training code on REALSR dataset using IMDN_AS arch?

Thanks.

I'm using your framework to train my own network. It seems that the process just saved every checkpoint, so, how did you find the best checkpoint.

how many epoch trained in IMDNx2 x3 x4？

as title , thank you

the location of adjusting learning rate should be at the end of train function

Shouldn't the Line 8 in train_IMDN.py be "import utils" instead of "from utils import utils"? And where does the function "utils.adjust_learning_rate()" used in Line 123 in train_IMDN.py first defined? Thank you.

hi, thanks for your work
could you tell me why you use rgb in 'as' model , and use ycbcr in 'x2/3/4/' model, are there any tricks?

Thanks for your excellent work.
But I hava a doublt about the results on x2 model.
I can't get the results as you report in the article, so should I use L2 loss after 1000 epochs?
Thanks!

Hello, how do I use your code to train pictures in real scenes (high-resolution images and low-resolution images are the same size)?

Hi, thank you so much for the source code. Besides, could you please provide the method for calculating memory？

您好当我使用IMDN_AS时，一旦改变upscale就会报错，是不可以更改这个参数吗？size mismatch for upsampler.0.weight: copying a param with shape torch.Size([48, 64, 3, 3]) from checkpoint, the shape in current model is torch.Size([12, 64, 3, 3]).
size mismatch for upsampler.0.bias: copying a param with shape torch.Size([48]) from checkpoint,

你好，为什么训练时候验证的psnr和test_imdn得出的psnr在同一个测试集上值是不同的呢

Thanks for your excellent works.
I have a question. I tested my images(3040x3040 or 6080x3040) with EDSR, little change in gpu memory . but, when i tested your IMDN with forward_chop, why the memory change so much?

Great job!
There is a question about the average inference time on BSD100, Urban100 and Manga109. In Table 6, for EDSR-baseilne, EDSR, RDN, and IMDN, the average inference time becomes less from BSD100 to Manga109. But the image size becomes larger from BSD100 to Manga109. That makes me confused. Did I miss something about this average inference time?

In addition, I used the "test_IMDN.py" script and tested the released "IMDN_x4.pth" on BSD100, Urban100 and Manga109 on 2080 Ti. The average inference times are 9.14ms, 14.37ms and 17.00ms. I think 2080 Ti should be faster than Titan Xp mentioned in your paper. Are there any wrong steps for me?

当运行train_IMDN文件时，set5的指标比运行test_IMDN时低了很多，两个文件用的都是train中最好的checkpoin