restormer的模型我这本地跑不了，但是这个模型一大优势就是可以直接处理整张图不掉那么多点。我这边基于MPRnet进行复现时发现与其整张图进行处理不如把图片切分成与训练一样的尺寸进行处理。
基于MPRnet的实验，我这patch=step=128处理时，涨点比直接处理明显；patch=128,step=64时比step=128的psnr要高一点，同理step=32也比step=64的处理要高一点点。patch与patch之间重合的区域是使用粗暴的平均取值的方式。step=patch/2的方式，华为给出了一个更佳的解决方法https://github.com/huawei-noah/Pretrained-IPT/issues/18，缺点就是step不能改小了。
我想问问，如果基于你新训练的restormer模型去切分图片进行处理的话，会不会还能再涨点呢？

Thanks for your contribution! There is problem when training,

File "/root/autodl-tmp/Restormer-master/utils.py", line 103, in __getitem__
    image_name = os.path.basename(self.rain_images[idx % self.num])
ZeroDivisionError: integer division or modulo by zero

Hi, very good work. But I have a small problem, my computer is a single graphics card Nvidia 3080, both batch_size and patch_size are smaller than the minimum size you proposed, but after 156000milestones, both batch_size and patch_size become bigger, then I get an error on my side——RuntimeError : pixel_unshuffle expects height to be divisible by downscale_factor, but input.size(-2)=49 is not divisible by 2, how do I fix this?

batch_size=16, 10, 8, 4, 2, 2
patch_size=32, 40, 98, 64, 80, 96

您好，看到您非常简洁清楚的复现代码，获益匪浅，但是我在复现您的代码时遇到一些问题，想向您请教，我在进行模型训练时，loss始终维持在0.175左右不再下降，psnr和ssim也分别保持在15和0.45左右不提高，想请教您这是什么原因如何改进，我的相关参数设置如下batch size=[32,20,16,8,4,4]，patch size=[64,80,96,128,160,192]，milestone=[9200,15600,20400,24000,27600]，num_iter=30000