Testing

[simoneVU]

Hi,

I was wondering when you plan to release the model weights and update the readme for model inference. That would be very interesting! Great paper btw!

Cheers, Simone

[zhengchen1999]

Thanks for your interest in our work. We plan to release training and testing code soon (probably in September, after the NeurIPS23 results are announced).

[simoneVU]

Thank you very much for the reply! looking forward!

[macvincent]

Hello, just following up on this to find out when the code will be released.

[zhengchen1999]

Dear @macvincent @simoneVU, we have released training&testing codes and pre-trained models.

[macvincent]

Awesome. Thank you @zhengchen1999. This is awesome work.

[TenMiss]

May I ask how to solve this problem：ModuleNotFoundError: No module named 'basicsr'

[zhengchen1999]

I think the reason for your error is that Installation are not prepared.

[DaddyWesker]

So, where I can find pretrained models? Here lies files DAT....pth. but code requires experiments/pretrained_models/HI_Diff_RealBlur_J/net_le_dm.pth experiments/pretrained_models/HI_Diff_RealBlur_J/net_g.pth experiments/pretrained_models/HI_Diff_RealBlur_J/net_d.pth

Moreover, it is unclear how to infer model on a custom image. Can you make infer script for custom input image?

[zhengchen1999]

The pretrained models are provided here.

To test on your own dataset, set yml: dataroot_gt and dataroot_lq to the corresponding directories (both should be the same).

[DaddyWesker]

THanks!

Resulting deblurred image should be there? HI-Diff/results/test_HI_Diff_RealBlur_R/visualization (i'm using RealBlur_R model and yml accordingly)

[zhengchen1999]

In the RealBlur_R folder (HI-Diff/results/test_HI_Diff_RealBlur_R/visualization/RealBlur_R). Note that the output results are saved only if 'save_img: True' is set in the yml.

[DaddyWesker]

This value is set to true. Well, I've ran this model on my img and it doesn't work i guess. Input:

Output:

Maybe I'm doing something wrong. I was thinking on debluring this image but wooden chair is still defocused... And it seems that nothing changed at all.

[zhengchen1999]

The test should be fine. However, I think your image is defocus-blur, which is inconsistent with our trained model. I suggest you use other defocus-deblur methods such as (Restormer). Of course, you can also try the model trained with GoPro (i.e., GoPro.yml).

[DaddyWesker]

I've tried restormer and it doesn't help either. I'll try GoPro variant, thanks. I've been struggling to find any algorithm to beat that defocus blur but everything failed so far. Thank you anyway.

[sunzhe09]

我也遇到同样的问题，感觉有一定的去模糊效果，但是处理后还是有一些模糊，不知道需要调整什么参数来优化

[alalnjcno]

请问test的时候 用GoPro的时候只有这一个数据集 input和target里面分别放什么呀 放一样的吗

[zhengchen1999]

是的，放一样的，结果在results文件夹中。

[alalnjcno]

你好，请问处理后的图象在哪里看呢

[zhengchen1999]

在results文件夹，yml文件中name对应文件夹，visualization文件夹中。

[alalnjcno]

你好 我使用HI_Diff_GoPro模型 进行测试RealBlur_J 显示 是我哪里出错了吗

[zhengchen1999]

不好意思，我没有遇到过这种情况。请问你的RealBlur_J结果是如何生成的。如果是通过python test.py -opt options/test/GoPro.yml测试得到，那相应的yml文件有修改过吗？

[alalnjcno]

不好意思，我没有遇到过这种情况。请问你的RealBlur_J结果是如何生成的。如果是通过python test.py -opt options/test/GoPro.yml测试得到，那相应的yml文件有修改过吗？

只测试一个数据集 就把GoPro.yml中多余的删掉了 然后属于以下两条指令python test.py -opt options/test/GoPro.yml python evaluate_realblur.py --dataset RealBlur_J --dir results/test_HI_Diff_GoPro
我在3080上进行测试的

[zhengchen1999]

你好，我在3090上重新测试了结果： For RealBlur_J dataset PSNR: 29.148628 SSIM: 0.889761 结果正常。因此，如果你没有修改过代码的话，可能是数据集出现问题了。你可以尝试下载一份新的数据集。此外，考虑到你的结果性能很高，会不会是用了target作为input数据（当然，你的yml中并没有设置错路径，这只是我的猜测。

[alalnjcno]

你好，我在3090上重新测试了结果： For RealBlur_J dataset PSNR: 29.148628 SSIM: 0.889761 结果正常。因此，如果你没有修改过代码的话，可能是数据集出现问题了。你可以尝试下载一份新的数据集。此外，考虑到你的结果性能很高，会不会是用了target作为input数据（当然，你的yml中并没有设置错路径，这只是我的猜测。

应该是数据集的问题，您的那个原数据集 是两个文件夹 分别包含清晰与模糊的图像对吗？我下载的好像只有一个文件夹里面是模糊的，还有一个为叫__MACOSX 的文件夹 里面图象损坏

[zhengchen1999]

__MACOSX请忽略，这是mac压缩造成的额外数据。 另外，请问你下载的数据集链接是哪个？我下载的结果是正常的，位置：https://github.com/zhengchen1999/HI-Diff?tab=readme-ov-file#datasets

PS: 只有"模糊"图像的，是否是我提供的视觉结果，也就是我复原出来的结果，位置：https://github.com/zhengchen1999/HI-Diff?tab=readme-ov-file#models

[alalnjcno]

__MACOSX请忽略，这是mac压缩造成的额外数据。 另外，请问你下载的数据集链接是哪个？我下载的结果是正常的，位置：https://github.com/zhengchen1999/HI-Diff?tab=readme-ov-file#datasets

PS: 只有"模糊"图像的，是否是我提供的视觉结果，也就是我复原出来的结果，位置：https://github.com/zhengchen1999/HI-Diff?tab=readme-ov-file#models

谢谢您 是我数据集下载错了 。还有想问下您在训练过程中 收敛时两个损失函数数值和 测试集上测试的INFO: Validation ValSet 上输出的PSNR大约是多少呀？我在一张卡上训练 训练结果在完整数据集上测试26左右

[zhengchen1999]

S1: l_pix: 1.5e-2, Validation: 33 dB S2: l_pix: 1.0e-2 l_pix_diff: 4.5e-1, Validation: 33 dB