idkiro / gunet Goto Github PK

Whether the author has a trained model

Hi

Thanks for the interesting work. I saw that there is a ./models/baselines folder in the repo. The paper also mentioned, " All models are trained using their original strategies, and we replicate the best results reported in the previous works". I am wondering if that means you put the models together in this repo and replicated the baselines based on their original training strategies for RESIDE-IN and RESIDE-OUT. And used the same training strategy for RS-Haze if the original work did not use this dataset?

==> Using SyncBN because of too small norm-batch-size.
==> Start training, current model name: gunet_t
0%| | 0/501 [00:00<?, ?it/s]

Is it necessary to implement the abstract method '_get_lr' in the base class in the 'scheduler.py' file?

Thank you very much for your sharing!
The following error occurred when I entered an image size of 412*548:
in_feats = torch.cat(in_feats, dim=1)
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 102 but got size 103 for tensor number 1 in the list.

What is the reason for this problem？How should I predict dehaze images?
Looking forward to your reply！

Why downSample is using 3*3 conv in paper framework, but i don't see kernel=3 in code?

请问在batchsize较小时，frozen bn训练提升大吗？ 如果batchsize较大，可不可以不用frozen bn

Hi @IDKiro,

Thanks for your impressive work. Recently we are going to do some related work. Hence I am curious about the training resources issue.

May I kindly ask how many GPUs are used when training such a model? And What kind of GPUs? And how long it takes for the training?

Best regards,

When I training, I meet a trouble:

at models/gunet.py : out=self.Wv(X) * self.Wg(X)

when X'size is torch.Size([16, 192, 2, 2])， dim=192，kernel_size=5,
self.Wg is wrong.

Could you have some method to solve this trouble?

self.Wv = nn.Sequential(
nn.Conv2d(dim, dim, 1),
nn.Conv2d(dim, dim, kernel_size=kernel_size, padding=kernel_size//2, groups=dim, padding_mode='reflect')
)

在configs中,您将ITS的 "valid_mode"设置为 "test", OTS的 "valid_mode"设置为"valid". 这似乎会导致使用OTS数据集训练模型时val_dataset中图片的大小为256*256. 而在test.py中,您似乎使用原始尺寸的图片来计算PSNR和SSIM.这是否会导致train.py中输出的PSNR和test.py中输出的PSNR不同?

您好，想问一下您 这下边的这个报错是什么问题呢？
File "/data1/lyl/gUNet-main/utils/common.py", line 52, in read_img
return img[:, :, ::-1]
TypeError: 'NoneType' object is not subscriptable
非常希望得到您的回复，感谢

请问作者使用Haze-4k数据集训练模型时用了多少张显卡，多长时间？
期待您的答复

运行train.py出现错误：ValueError: With replacement=False, num_samples should not be specified, since a random permute will be performed.
想咨询一下如何解决。

Traceback (most recent call last):
File "D:/Edge_Downloade/gUNet-main/gUNet-main/train.py", line 213, in
main()
File "D:/Edge_Downloade/gUNet-main/gUNet-main/train.py", line 172, in main
sampler=DistributedSampler(val_dataset, shuffle=False), # comment this line for more accurate validation
File "D:\anaconda\envs\Other\lib\site-packages\torch\utils\data\distributed.py", line 66, in init
num_replicas = dist.get_world_size()
File "D:\anaconda\envs\Other\lib\site-packages\torch\distributed\distributed_c10d.py", line 867, in get_world_size
return _get_group_size(group)
File "D:\anaconda\envs\Other\lib\site-packages\torch\distributed\distributed_c10d.py", line 325, in _get_group_size
default_pg = _get_default_group()
File "D:\anaconda\envs\Other\lib\site-packages\torch\distributed\distributed_c10d.py", line 429, in _get_default_group
raise RuntimeError(
RuntimeError: Default process group has not been initialized, please make sure to call init_process_group.