JingyunLiang / SwinIR

SwinIR: Image Restoration Using Swin Transformer (official repository)

Home Page:https://arxiv.org/abs/2108.10257

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

关于PSNR和SSIM没有收敛到原论文中的性能

lyongo opened this issue · comments

commented

作者你好,

谢谢你所做的非常不错的工作,我阅读了SwinIR论文,并且star了此仓库。在我使用你们提供的预训练模型在Set5数据集上测试 和在DIV2K和Flickr2K数据集训练Class Imager(x2)时,发现PSNR值没有达到原论文中的值,在这里请教下是否因为我的超参数设置的问题还是有些训练的trick。

使用官方的预训练模型(001_classicalSR_DF2K_s64w8_SwinIR-M_x2.pth),在Set5上测试:
image

Average PSNR: 36.21 dB;

使用https://github.com/cszn/KAIR 中提供的训练代码 在DIV2K和Flickr2K数据集训练Class Imager(x2) 时:
image

Average PSNR 收敛到36.15dB,没达到论文中的性能。

谢谢!

  1. We compare the Y-channel PSNR for image SR, the pretrained model achieves the same results as in Table 1 of the paper.

  2. For your own training, do you use the same batch size and other settings? It is also possible that different runs may generate slightly different results (<0.06dB).

  3. Lastly, could use the mian_test_swinir.py to test your trained model on other datasets? What are their PSNR?

commented
  1. We compare the Y-channel PSNR for image SR, the pretrained model achieves the same results as in Table 1 of the paper.
  2. For your own training, do you use the same batch size and other settings? It is also possible that different runs may generate slightly different results (<0.06dB).
  3. Lastly, could use the mian_test_swinir.py to test your trained model on other datasets? What are their PSNR?

谢谢您的回复 ,

  1. 在SwinIR 论文的Table 2 中与SwinIR对比的Classical image SR 方法的PSNR的指标使用的都是 Y-channel PSNR 吗?

  2. 如何使用自己训练的SwinIR模型在mian_test_swinir.py中测试呢?

1, Yes. We pointed it out in experimental setup.

2, Just use your own trained model path

ezoic increase your site revenue