yjn870/rdn-pytorch

PyTorch implementation of Residual Dense Network for Image Super-Resolution (CVPR 2018)

rdn-pytorch's Issues

the problem of RDB

I find the code you write about RDB is wrong: you lost the connection about DenseLayer

[Feature Request]: Add LICENSE

Thanks for your work! Could you please add a license to the project for a proper reuse of your code?

预训练权重文件

你好，请问能给出百度云的预训练权重文件的链接吗？ 现在这个打不开~

lr decay

param_group['lr'] = args.lr * (0.1 ** (epoch // int(args.num_epochs * 0.8)))

In the paper : The learning rate is initialized to 10−4 for all layers and decreases half for every 200 epochs.

I don't understand what you are doing with the learning rate, can you explain it, thanks a lot

The value of the PSNR in this program seems to be different from other programs

Hello!
Are the PSNR results recovered from the model high? I tested with the pictures from the set5 recovered by the model in other programs that calculate PSNR, and find that the values are not as high as in your programs, (and the programs I used to test PSNR has been verified in Set5 Bicubic's case is 33.67, it seems that there is no problem), can you explain this phenomenon? And if you can provide SSIM program is even more perfect.
Thanks!

why use convert_rgb_to_y

In the 3.4. Implementation Details section of the original paper it is stated that the network applies grayscale images.And there is no explicit mention of pre-processing in the paper.Simple cropping is also done in the matlab code in the github repository given in the paper.

convert_rgb_to_ycbcr

Hello, have you forgotten this function “convert_rgb_to_ycbcr” in "util.py"? I can't find the declaration.