zhilin007 / ffa-net Goto Github PK

Hi,

Thanks for open-sourcing your work. I was trying to write a TensorFlow implementation of your model, and ran into the following issue.

In FFA.py, you have the code below at L96:

w=w.view(-1,self.gps,self.dim)[:,:,:,None,None]
out=w[:,0,::]*res1+w[:,1,::]*res2+w[:,2,::]*res3

Questions:

1. After the first line, tensor w will have shape [B, gps, dim, 1, 1], where B is the batch size. Is that correct?
2. Subsequently, in the second line, w[:, 0, ::] has shape [B, dim, 1, 1]. Correct?
3. But res1 has shape [B, H, W, dim], where H and W are the spatial dimensions. How could you multiply w[:, 0, ::] and res1 which have incompatible shapes?

Sorry if this is trivial - I normally write TensorFlow and am quite unfamiliar with the specifics about PyTorch. Thanks!

我不知道是上传的代码版本有问题还是别的原因，但是DataLoader确实有问题，它没有返回hazy and gt， 而是hazy和正则化的hazy(名字为clear)。 这就是为什么指标很好，但是视觉上看没有效果的原因：因为dataloader的问题，模型拟合的目标一开始就是错的。
Error in class dataloader , it should return hazy image and clear image, but dataloader returns normalized hazy image and hazy image. This is why the metrics are good, but visually ineffective: because of the dataloader, the goal of the model fitting was wrong in the first place.

Code below looks normal but "clear" is actually hazy image.

clear_name = img    
clear = Image.open(os.path.join(self.clear_dir, clear_name))  

Advised Solution:

(__init__)
self.clear_imgs_dir = os.listdir(os.path.join(path, 'GT'))
self.clear_imgs = [os.path.join(path, 'GT', img) for img in self.clear_imgs_dir]                  
（__getitem__）
clear = Image.open(self.clear_imgs[index])

Hello, how to print the channel-wise and pixel-wise feature weights map of the Group Structure output? Can it be printed in training the dataset or after training the dataset? Thank you very much!

这论文竟然对真实有雾图几乎没用

I believe that if you use the global residual connection, you need to normalize both of them.
Unless, the residual term has to learn much more complicated space.

I used your code to train and test on RESIDE V0 dataset but I got the result 0.84/20.57 (SSIM/PSNR) for indoor task. Was I wrong at any step?

请问如何测试处理单张图片的时间？

![haze3](https://user-images.githubusercontent.com/56646894/206660018-8ed5eaf8-5f

39-4baa-9f53-45355868320a.jpg)

It's not working. Did I write the wrong code?
question.txt

I'm not chinese and can't sign up to the Baidu homepage if I don't have phone number in China...
Could you upload the trained model in other platform like Google drive or Dropbox or somewhere?

We have asked AAAI to recall your paper, as it's impossible to achieve the same perfmance in your paper.
Vrevision it if there is something wrong. Or we will let your college and mentor informed.
Academic fraud is shameful.

Hi, thank you for releasing the code and pre-trained models.

When I test the pre-trained model ('its_train_ffa_3_19.pk') downloaded from the released website, I got
"loading state_dict for DataParallel:
Unexpected key(s) in state_dict: "module.g1.la.gamma", "module.g1.last.weight", "module.g1.last.bias", "module.g2.la.gamma", "module.g2.last.weight", "module.g2.last.bias", "module.g3.la.gamma", "module.g3.last.weight", "module.g3.last.bias"."

Do you know why?

Thanks a lot!

Regards.

Why is there a big difference between the same picture using the PSNR and SSIM methods here and the method provided on the RESIDE official website? What caused it？

D:\python\envs\pytorch\python.exe E:/FFA-Net-master/net/main.py
Namespace(blocks=20, bs=16, crop=False, crop_size=240, device='cpu', eval_step=5000, gps=3, lr=0.0001, model_dir='./trained_models/its_train_ffa_3_20.pk', net='ffa', no_lr_sche=False, perloss=False, resume=True, steps=100000, testset='its_test', trainset='its_train')
model_dir: ./trained_models/its_train_ffa_3_20.pk
16
E:\FFA-Net-master\net
crop size whole_img
crop size whole img
crop size whole_img
crop size whole img
log_dir : logs/its_train_ffa_3_20
model_name: its_train_ffa_3_20
train from scratch ***
Traceback (most recent call last):
File "E:/FFA-Net-master/net/main.py", line 156, in
train(net,loader_train,loader_test,optimizer,criterion)
File "E:/FFA-Net-master/net/main.py", line 62, in train
x,y=next(iter(loader_train))
File "D:\python\envs\pytorch\lib\site-packages\torch\utils\data\dataloader.py", line 346, in next
data = self._dataset_fetcher.fetch(index) # may raise StopIteration
File "D:\python\envs\pytorch\lib\site-packages\torch\utils\data_utils\fetch.py", line 47, in fetch
return self.collate_fn(data)
File "D:\python\envs\pytorch\lib\site-packages\torch\utils\data_utils\collate.py", line 79, in default_collate
return [default_collate(samples) for samples in transposed]
File "D:\python\envs\pytorch\lib\site-packages\torch\utils\data_utils\collate.py", line 79, in
return [default_collate(samples) for samples in transposed]
File "D:\python\envs\pytorch\lib\site-packages\torch\utils\data_utils\collate.py", line 55, in default_collate
return torch.stack(batch, 0, out=out)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 460 and 407 in dimension 2 at C:\w\1\s\tmp_conda_3.6_081743\conda\conda-bld\pytorch_1572941935551\work\aten\src\TH/generic/THTensor.cpp:689

Process finished with exit code 1

请问在室外场景的训练集、测试集中，RESIDE数据集给出的室外场景OTS训练集和室外场景测试集SOTS是有重叠的，您是如何解决的这个问题？

I have use you trained models to test on real hazy image (from google) but it look not good. Have I done something wrong?
(left: hazy imge - right: prediction using ots model)

我用作者提供的代码在我的设备上运行main.py,报错内存不足，只有当我把batchsize设置为1时才能运行成功，但是我用的时TeslaV100 32G显卡，作者用的是2080Ti，作者设置的batchsize是16，为什么我的会内存不足呢，有同学遇到这个问题吗

Thanks for sharing your great work!

I am a collector of great models. I basically collect models in the repositories below with the consent of the model authors, convert them from PyTorch to other frameworks such as TensorFlow or ONNX, and publish them.
https://github.com/PINTO0309/PINTO_model_zoo

By the way, I have already succeeded in converting your model.
PyTorch -> ONNX, TFLite, CoreML, TFJS, TF-TRT, OpenVINO, Myriad Inference Engine Blob

However, I am hesitant to publish models whose license is not clear, no matter how great they are. If possible, could you please add a LICENSE file? I am assuming that I will be quoting under the license you guys specify. For example, Apache2.0, MIT, and GPLv3, etc...

Thank you. 😄

提示内存不够，我在任务管理器中看了内存条32g的确是将近满了，CPU90%以上，我有两个GPU,GPU0是Intel(R) HD Graphics 630约30%，GPU1是NVIDIA GeForce GTX 1050 Ti2%基本没变化，我感觉我NVIDIA显卡没派上用场，你的电脑配置是什么样的跑成功了吗，你是使用什么软件运行这个代码的？

hi，The test input is the pre-processing result after normalization（ tfs.Normalize(mean=[0.64, 0.6, 0.58],std=[0.14,0.15, 0.152])）, while the output needs to be de-normalized to get the RGB image with the output pixel value between 0 and 255. Could you please tell me how to de-normalize it?

https://pan.baidu.com/error/404.html

If I use my own dataset for training, do I not need the model trained by the author before?
Thanks very much！

I am sorry to disturb you again.
"The number of Group Structure G is 3. In each Group Structure, we set the Basic Block Structure number as B = 19."
Why G is 3, B is 19?
Thank you very much!

Hello, I want to ask how to visualize the attention map.
Thanks

code in main.py 19th line
why mean=[0.45627367, 0.38360259, 0.36426624]#clear?
how to calculate mean?
thank you!