yjn870 / fsrcnn-pytorch Goto Github PK

hello，当我使用仓库代码进行训练时，在set5上测试精度只有35.12，达不到仓库里的37.12，请问您的训练策略是什么？

nn.PReLU() shouldn't set x or s in Sequential, for example, from my code, I think nn.PReLU(x) is false in Sequential(It will cause bug torch.cuda.FloatTensor is not a Module subclass). We should change it nn.PReLU(), it seems right in Sequential

When testing, why are the input and output image sizes the same regardless of the scale? And why is the test input a GT image?

hi, I checked your image transformation part in file utils.py
the formula for rgb to ycbcr image is a little different from Wikipedia:
yours:
y = 16. + (64.738 * img[..., 0] + 129.057 * img[..., 1] + 25.064 * img[..., 2]) / 256.
wiki:
y = 16. + (65.738 * img[..., 0] + 129.057 * img[..., 1] + 25.064 * img[..., 2]) / 256.

and actually in all of your implementations, you use 64.738.
i tested both and seems this minor difference does not affect the prediction accuracy...
but still, want to point out this and make this work better.

你好，感谢您提供相关代码。
在本仓库运行时，使用您提供的预训练模型，测试butterfly 2/3倍scale时，CPU耗时700ms，GPU(NVIDIA-SMI 515.65.01 Driver Version: 515.65.01 CUDA Version: 11.7)耗时400ms，
与论文中的结论27fps有差异，是什么原因呢

Hello, I want to enlarge the output image by 4 times. I set the stride in ConvTranspose2d to 4, but this will give an error, which will indicate that the dimensions of tensor a and tensor b are different. Can you tell me why this is happening, What will be done to magnify the output image by 4 times for me?