li-chongyi / dehamer Goto Github PK

How many GPUs are used during training?
In the paper, it is mentioned a single V100 is used for implement, wondering the number of GPUs used for training.

您好，我想请问一下，您的模型在真实雾天场景下有测试过吗？我这边测试真实场景并没有变化

Dear Author, where did you get the results on Dense-Haze dataset? I did not find the results of other methods in their papers.

And the results of other methods on Dense-haze are different from that in paper 'A novel encoder-decoder network with guided transmission map for single image dehazing'.

Thanks for your explanation in advance.

hello, I think google drive link for ITS-data doesn't work now.
can I get drive-link for dataset?

I am unable to reproduce the results on NH-HAZE dataset by training from scratch, can anyone give me some advice please?

请问val_data和val_data_train这两个py文件哪个是用来测试集的数据读取呢？两个.py文件有什么区别呢？

作者您好，您有试过用Dehamer模型在CPU上进行推理吗？我用NH的模型测试NH数据集的图片，在GPU上推理正常，但是如果用CPU推理出来，保存的图片却是全黑的，您知道这是为什么吗？

when compared with 608X448 Sots GT (indoor), the results are suprisingly low, including the offical results released, below 17psnr average

请问你是否能尽快上传一下对比实验及你网络在Dense-Haze和NH-haze这两个测试数据集的output

Hi,
I’m currently trying to train the model on Nh-Haze dataset, and I found that in the Section 4.1 of the paper, you have mentioned that you gradually enlarge the size of the image patch from 128×128 to the full size in the training process, but I didn’t find it in the code. I’m wondering about the enlarging details, including when to enlarge the size and how large the size will be?
Thank you so much!!

权重有swin_2，你代码里没有啊，请你检查一下

Hi

Thanks for the exciting work. I am trying to reproduce your results by training from scratch. I found some missing details in the paper and the training settings in the train.sh are all the same for different datasets. I am wondering what is the exact training setting you used for outdoor subset of the RESIDE dataset?

Two functions "A_from_BCP()" and "T_min_from_A()" were missed, could you update them, thanks.

1，我在真实数据上进行测试，效果十分不理想。这种虽然PSNR和SSIM高，但效果未必好，如何解决呢？如果只追求更高的PSNR和SSIM，也只能说其回归性能更逼近真值，却未必能说解决真实问题。域适应的方法不知道作者是否有所考虑？
2，使用合成数据训练，倘若真值本就有薄雾，在真值上再加雾。然后训练模型，训练的模型能否突破原有真值的上界，将真值中本就存在的雾去掉呢？我觉得这是不可能的，因为现在的训练方法只是去逼近真值，无法超越真值。

In paper 4.1 Experimental Settings, I found that when training the network the image patch is from small to full size, but I do not see the corresponding part in the code. Will this affect the training of the model? And the batch size is 2, which maybe is too tiny for V100 GPU. Why did you set such a small value?