Results 8 issues of jie

## 🌱 Describe your Feature Request nn.PixelUnshuffle in pytorch 1.9 is not supported.

feature request
pytorch
missing layer type

make failure [ 80%] Linking CXX shared library x86_64/lib/libsysDetectSpeed.so /usr/bin/ld: cannot find -lopencv_dep_cudart /usr/bin/ld: cannot find -lopencv_dep_nppial /usr/bin/ld: cannot find -lopencv_dep_nppicc /usr/bin/ld: cannot find -lopencv_dep_nppicom /usr/bin/ld: cannot find -lopencv_dep_nppidei /usr/bin/ld:...

Hi, I use the preprocessed CelebA dataset that you provided to train, but the attention masks become 1 after several iters. I didn't change the parameters in the code. `...

I train the model from scratch and use the default parameters,but the loss remains too large(1e4) after 40000 iters.

hi, thanks for your great work. I have a problem when i run lowlight_test.py. 刚开始执行没有问题,测试一定数量的图片之后,提示显存不足,暂时没发现什么问题。 ` data/test_data/DICM/32.jpg 0.001220703125 data/test_data/LIME/6.bmp 0.0012459754943847656 data/test_data/LIME/8.bmp 0.0011720657348632812 data/test_data/LIME/4.bmp 0.0012981891632080078 data/test_data/LIME/7.bmp 0.0016331672668457031 data/test_data/LIME/1.bmp 0.0012412071228027344 data/test_data/LIME/2.bmp 0.0011746883392333984...

hello,我使用pytorch版的测试的,把两张一样的图组成batch 2的tensor,调用接口进行边缘检测,输出的两个结果不一样,和单独测试一张图的结果也不一样。

模型用的RealworldSR-DiffIRS2-GANx2.pth, 在DiffIR-RealSR目录下执行: ` python inference_diffir.py --im_path ../../dataset/test/ --res_path results/ --model_path experiments/DiffIRS2-GANx2/RealworldSR-DiffIRS2-GANx2.pth --scale 2 ` 原图: ![9-4](https://github.com/Zj-BinXia/DiffIR/assets/22720234/97974fa9-58c0-42ee-bc14-a8af4f906cad) 超分后: ![9-4_out](https://github.com/Zj-BinXia/DiffIR/assets/22720234/505f7874-2b3f-4fdf-b3bb-dddf59cc2f6f)

对于ppocrv4系列模型,paddle训练模型转换成pytorch后,模型比paddle的infer模型大很多。 比如,ch_PP-OCRv4_rec_train模型转成pytorch后26M左右,pt转成onnx后25M左右,但是paddle的infer模型只有10M左右,都是fp32。 检测模型也是类似,ch_PP-OCRv4_det_train模型转成pytorch后14M左右,但是paddle的infer模型只有5M左右,也都是fp32。