jimmychame/sadnet

Pytorch code for "Spatial-Adaptive Network for Single Image Denoising"

sadnet's Issues

How is this code compatible with pytorch1.7.0?

I tried to run the code in pytorch1.7.0 but found some mistake. So could you tell me how to tackle the version-imcompatible problem? thx.

import deform_conv_cuda

File "/workspace/divya/Denoising/SADNet/dcn/deform_conv.py", line 10, in
from . import deform_conv_cuda
ImportError: /workspace/divya/Denoising/SADNet/dcn/deform_conv_cuda.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN3c1019UndefinedTensorImpl10_singletonE

Kindly help

Log Files from Training

Thank you for your awesome code!

I am hoping you might open-source the log files you have from training. Maybe the training and validation loss as a function of epoch
(and/or batch) with an estimate of the runtime?

CUDA

请问电脑没有独立显卡，只有显卡适配器：spacedesk Graphics Adapter、Intel(R) UHD Graphics 770，可以运行此代码吗

cuda available issue

I've already finished the install step.
When I try to run the test.py, I met the error "There are not available cuda devices!"
But I've already had the cuda and cudnn in my environment which can be used in other times, and I add this line "print(torch.cuda.is_available())" in the main func of the test.py, it returns True to me.
But when it was the same 'torch.cuda.is_available()' in the evaluate_net() func , it turns into false.
Could you please help me to solve this problem?

A kindly reminder

Here is a kindly reminder. It seems that the result on sidd validation set is wrong duo to the image value range to computing the PSNR. As I tested the model, the PSNR would be 39.534db when the range is [0,1], but become 39.278db when the range is [0, 255]. There is a gap between two results. As far as I knowm, the range for the sidd bench mark to computing PSNR is [0, 255]. Many thanks.

论文问题请教

您好，我很有幸看了您的论文，写的非常好。但是我个人对于一些细节不太懂，麻烦请解答一下我的疑惑。万分感谢

我的理解是卷积是个局部操作，只能提取局部特征，而纹理和边缘属于高维的特征，需要更大的感受野才能提取。如果去噪网络层数过少就不易于高维特征的提取，进而出现过平滑。虽然随着网络的增加，感受野也增加，但是传统的卷积神经网络层数增加过多会导致计算花费的增加，因此，您提出新的结构在不增加层数的基础上增大卷积神经网路的感受野。（请问对不对？）
1、A traditional CNN can use only the features in local fixed-location neighborhoods, but these may be irrelevant or even exclusive to the current location. Due to their inability to adapt to textures and edges.（我对这句话不是很明白 even exclusive to the current location？什么意思）

python setup.py develop problem

root@f23dbc103f93:/workspace/SADNet/dcn# python setup.py develop
running develop
running egg_info
writing deform_conv.egg-info/PKG-INFO
writing dependency_links to deform_conv.egg-info/dependency_links.txt
writing top-level names to deform_conv.egg-info/top_level.txt
reading manifest file 'deform_conv.egg-info/SOURCES.txt'
writing manifest file 'deform_conv.egg-info/SOURCES.txt'
running build_ext
building 'deform_conv_cuda' extension
creating /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8
creating /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/src
Emitting ninja build file /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/build.ninja...
Compiling objects...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
1.10.1
creating build/lib.linux-x86_64-3.8
g++ -pthread -shared -B /opt/conda/compiler_compat -L/opt/conda/lib -Wl,-rpath=/opt/conda/lib -Wl,--no-as-needed -Wl,--sysroot=/ /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/src/deform_conv_cuda.o /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/src/deform_conv_cuda_kernel.o -L/opt/conda/lib/python3.8/site-packages/torch/lib -L/usr/local/cuda/lib64 -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-3.8/deform_conv_cuda.cpython-38-x86_64-linux-gnu.so
g++: error: /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/src/deform_conv_cuda.o: No such file or directory
g++: error: /workspace/SADNet/dcn/build/temp.linux-x86_64-3.8/src/deform_conv_cuda_kernel.o: No such file or directory
error: command 'g++' failed with exit status 1

environment

root@f23dbc103f93:/workspace/SADNet/dcn# g++ -v
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/9/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none:hsa
OFFLOAD_TARGET_DEFAULT=1
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu 9.3.0-17ubuntu1~20.04' --with-bugurl=file:///usr/share/doc/gcc-9/README.Bugs --enable-languages=c,ada,c++,go,brig,d,fortran,objc,obj-c++,gm2 --prefix=/usr --with-gcc-major-version-only --program-suffix=-9 --program-prefix=x86_64-linux-gnu- --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --disable-vtable-verify --enable-plugin --enable-default-pie --with-system-zlib --with-target-system-zlib=auto --enable-objc-gc=auto --enable-multiarch --disable-werror --with-arch-32=i686 --with-abi=m64 --with-multilib-list=m32,m64,mx32 --enable-multilib --with-tune=generic --enable-offload-targets=nvptx-none=/build/gcc-9-HskZEa/gcc-9-9.3.0/debian/tmp-nvptx/usr,hsa --without-cuda-driver --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 9.3.0 (Ubuntu 9.3.0-17ubuntu1~20.04)

What is the different between ModulatedDeformConvPack2 and ModulatedDeformConvPack

Hello, would you tell me What is the difference between ModulatedDeformConvPack2 in dcn.deform_conv and ModulatedDeformConvPack in dcn.deform_conv? The code of them are same.

deform_conv_cuda import error

Hello, i develop setup.py under Windows system, and it results in a .pyd file. But I had an import error like follows:
from . import deform_conv_cuda
ImportError: DLL load failed
Do you know how to solve this problem?
Best regards.

initialization method during training

Hi, thanks for sharing! I wonder when training conv2d for deformable conv_offset_mask, can the default xaiver uniform initialization method work?
In the original paper of DCN, the initilization method is constant zeros. which can prevent loss from exploding.
And have you tried effects of different initialization methods for conv_offset_mask learning? Personally, I think initialization method matters since deformable convolution is quite sensitive.

Testing Dataset

Hi could you please provide the dataset for testing

ImportError: cannot import name 'deform_conv_cuda'

Traceback (most recent call last):
File "test.py", line 10, in
from model.init import make_model
File "/userhome/sjc/SADNet/model/init.py", line 1, in
from model.sadnet import SADNET
File "/userhome/sjc/SADNet/model/sadnet.py", line 6, in
from dcn.deform_conv import ModulatedDeformConvPack2 as DCN
File "/userhome/sjc/SADNet/dcn/init.py", line 1, in
from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv, ModulatedDeformConvPack,
File "/userhome/sjc/SADNet/dcn/deform_conv.py", line 10, in
from . import deform_conv_cuda
ImportError: cannot import name 'deform_conv_cuda'

Issue when I run test.py

Traceback (most recent call last):
File "/home/ztzhao/Documents/code/papers/SADNet/model/sadnet.py", line 7, in
from dcn.deform_conv import ModulatedDeformConvPack2 as DCN
File "/home/ztzhao/Documents/code/papers/SADNet/dcn/init.py", line 1, in
from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv, ModulatedDeformConvPack,
File "/home/ztzhao/Documents/code/papers/SADNet/dcn/deform_conv.py", line 10, in
from . import deform_conv_cuda
ImportError: cannot import name 'deform_conv_cuda' from 'dcn' (/home/ztzhao/Documents/code/papers/SADNet/dcn/init.py)

How to install DCNv2 module with CUDA 10.1?

I have Pytorch 1.1.0 & CUDA 10.1, how to install DCNv2 module with CUDA 10.1?
When I run "python setup.py develop", I can't from . import deform_conv_cuda
Looking forward to your reply!

Google Drive 的models 下载不了

您好，目前打不开Google Drive链接， 可以把models上传到百度网盘上吗？

Why RSABs are not used at encoder?

Hi, from ablation study in your paper, RSAB without offset transfer seems to achieve the most denoising performance. Since RSAB without offset transfer can be used at encoder, why not do that? Intuitively, performance could be much better.

How can I apply the SADNet in my own noised img?

is the SADNet universal to most of imgs except your testdata in the paper?