gewelsji / pns-net

2021-MICCAI-Progressively Normalized Self-Attention Network for Video Polyp Segmentation

Hey Welcome to Daniel's Homepage

  Updated News

• [24/August/2022] ‼️ We present a new task, video polyp segmentation (VPS), which has been accepted by Machine Intelligence Research (MIR). We release the first large-scale VPS dataset, termed SUN-SEG, containing 158,690 frames with densely-annotated labels. These labels can further support the development of medical colonoscopy diagnosis, localization, and their derivative tasks. For more details, please refer to our project page / technical report.

• [06/August/2022] ❗ Our paper about camouflaged object detection (COD) has been accepted by Machine Intelligence Research (MIR) journal. This is a simple but efficient baseline, DGNet, with a novel object gradient supervision for the COD task. Additionally, we construct a comprehensive COD benchmark with 20 competed approaches. Read our technical report for more details. We also implement our model via Jittor & PyTorch toolboxes.

  Real-time Stats

Languages and Tools

pns-net's Issues

数据集已经失效了

作者，您好！论文中的数据集链接失效了，能提供以下最新的连接吗？

想问一下怎么提高代码cuda使用版本

看到应该跟CUDAExtension之类的有关，请问应该修改，非常感谢

Training problem

您好,请问这个是怎么回事
/PNS-Net-main/lib/PNS/self_cuda_backend.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN3c105ErrorC1ENS_14SourceLocationERKS

Could you provide the training file?

I encounter a problem when I am trying to implement your work on my own(video shadow segmentation). The Normalized Self-attention block does not work as you do.

作者您好，问您一个问题

cd ./lib/PNS
python setup.py build develop
这一点安装是在什么地方呀？cd 找不到路径，谢谢！

训练技巧和数据集的分配

    您好，我最近有关注你们的工作，十分地出色！但是我遇到了一些问题，首先是我使用了你们的方法，无法达到你们的精度，我想问问您有没有使用一些tricks，还有您们的工作会再更新pytorch版本吗？下一个疑问是，对于您们的数据集的划分还是存在一些困惑，IVPS-TrainSet这个数据集是不是没有用到？验证集和测试集是用相同的吗？还是怎么去处理？
     期待您的回复，谢谢

ModuleNotFoundError: No module named 'self_cuda_backend'

您好，请问ModuleNotFoundError: No module named 'self_cuda_backend' 应该如何解决呢？

ModuleNotFoundError: No module named 'self_cuda_backend'

作者，您好，我已经按照您在Readme中的环境配置命令配置相应环境并运行了“python setup.py build develop”命令，运行期间没有报错，但运行完成后似乎没有生成“self_cuda_backend”文件，而且代码还是会报错，下图是我这边运行后的文件结构不知道是不是我这边运行错误，如果运行正常，会在哪里生成“self_cuda_backend”文件，麻烦您解答一下，感谢！

从头训练没有达到论文中的精度

作者您好，我用您的代码从头训练但是测试结果并没有达到论文中给的精度，尤其在CVC-ColonDB-300数据集上，精度会差将近10个百分点，请问这是怎么回事呢，方便邮箱沟通一下吗

Some questions about the NL module

Hello
I would like to ask why it needs to redefine the backward for NL modules, instead of writing the module and using the backward in pytorch, because of the "processing" operation?
I am not familiar with custom functions and I hope you can answer this for me.

缺少文件

ModuleNotFoundError: No module named 'self_cuda_backend'
工程需要 self_cuda_backend 文件吧，谢谢解惑！

文件缺失

没有 MyTrain_Pretrain.py 这个文件

执行setup.py报错问题

您好，打扰一下，想问一下我在执行setup.py时出现如图所示错误可能是什么原因呢，python3.6，cuda12.0，cudatoolkit=10.0

Thanks for the excellent job！

Thanks for the excellent job！
There is my questions:

1. I am confused about the file PNS-Finetune.pth that you provide in the "Training/Testing" part. Whether it is the final weights or the pre-trained weights?
2. Could you please explain the 'max' in the metrics 'maxDice', 'maxSpc' and 'maxIoU'？
3. I find something wrong in the file utils/dataloader.py, where the variate 'begin' is supposed to be set to 0 not 1 in class VideoDataset. The file MyTest_Finetune.py is the case.
   It would be very grateful if you could answer my questions.

无法复现论文精度

您好，我阅读了您在VPS任务上的工作，是非常优秀和solid的工作，但是我在使用您的代码复现论文结果的过程当中遇到了一些问题，在您给出的三个数据集上很难达到论文给出的指标性能（尤其是cvc-colon-300上maxIOU和maxDice指标相差将近8、9个点）。
我的训练方式参照预训练100epochs + 微调1个epoch的方式，超参的设置使用了config.py文件中给出的设置，并且在微调时候替换了pretrained的路径。

提示缺少cuda.h如何解决

在生成编译文件时提醒我缺少gcc，并且提示我缺少cuda.h,但是我的cuda是正常安装了10.1版本的，而且也是正常安装gcc的。如下所示，请问该如何解决？
(PNSNet) yao@yao-System-Product-Name:/home/PNS-Net-main/lib/PNS$ python setup.py build develop
running build
running build_ext
building 'self_cuda_backend' extension
gcc -pthread -B /home/yao/anaconda3/envs/PNSNet/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/yao/anaconda3/envs/PNSNet/lib/python3.6/site-packages/torch/include -I/home/yao/anaconda3/envs/PNSNet/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/yao/anaconda3/envs/PNSNet/lib/python3.6/site-packages/torch/include/TH -I/home/yao/anaconda3/envs/PNSNet/lib/python3.6/site-packages/torch/include/THC -I:/usr/local/cuda-10.1/include -I/home/yao/anaconda3/envs/PNSNet/include/python3.6m -c PNS_Module/sa_ext.cpp -o build/temp.linux-x86_64-3.6/PNS_Module/sa_ext.o -std=c++11 -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=self_cuda_backend -D_GLIBCXX_USE_CXX11_ABI=1
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++
In file included from PNS_Module/sa_ext.cpp:3:0:
PNS_Module/utils.h:6:10: fatal error: cuda.h: 没有那个文件或目录
#include <cuda.h>
^~~~~~~~
compilation terminated.
error: command 'gcc' failed with exit status 1

训练技巧提问

尊敬的作者您好，我在尝试复现您的论文PNS-Net网络，在您提供的数据集上进行预训练和微调，测试结果在Dataset:CVC-ColonDB-300数据集上没有达到论文中给的精度，三个训练集的评价指标结果如下。

(Dataset:CVC-ClinicDB-612-Test) meanDic:0.774;meanIoU:0.696;wFm:0.765;Sm:0.864;meanEm:0.857;MAE:0.051;maxEm:0.884;maxDice:0.803;maxIoU:0.729;meanSen:0.757;maxSen:0.992;meanSpe:0.975;maxSpe:0.985.
(Dataset:CVC-ClinicDB-612-Valid) meanDic:0.843;meanIoU:0.756;wFm:0.824;Sm:0.922;meanEm:0.936;MAE:0.013;maxEm:0.966;maxDice:0.883;maxIoU:0.810;meanSen:0.862;maxSen:0.991;meanSpe:0.968;maxSpe:0.978.
(Dataset:CVC-ColonDB-300) meanDic:0.654;meanIoU:0.551;wFm:0.617;Sm:0.838;meanEm:0.797;MAE:0.025;maxEm:0.835;maxDice:0.693;maxIoU:0.590;meanSen:0.764;maxSen:1.000;meanSpe:0.978;maxSpe:0.990.

麻烦请教一下您，训练模型时有什么其他的技巧吗？另外，预训练和微调之后的网络模型测试评价指标差距不是很大，可以增加微调epoch或者有其他方法增强微调网络的性能吗？多有打扰。

Pretrain【epoch100】

(Dataset:CVC-ClinicDB-612-Test) meanDic:0.771;meanIoU:0.693;wFm:0.770;Sm:0.856;meanEm:0.852;MAE:0.050;maxEm:0.888;maxDice:0.811;maxIoU:0.736;meanSen:0.734;maxSen:0.992;meanSpe:0.974;maxSpe:0.983.

Finetune【epoch1】

(Dataset:CVC-ClinicDB-612-Test) meanDic:0.774;meanIoU:0.696;wFm:0.765;Sm:0.864;meanEm:0.857;MAE:0.051;maxEm:0.884;maxDice:0.803;maxIoU:0.729;meanSen:0.757;maxSen:0.992;meanSpe:0.975;maxSpe:0.985.

Training Code?

Where is the MyTrain_Pretrain.py and MyTrain_finetune.py file?

关于训练模型的问题

Remember to configure the pretrain_state_dict in config.py for different training stages 这个是什么文件？可以提供一下吗？