hi，ZhouHuang：
Thanks for your sharing, but when I tried to run test.py, it raise a error which told me"ValueError: ImageIO does not generally support reading folders. Limited support may be available via specific plugins. Specify the plugin explicitly using the plugin kwarg, e.g. plugin='DICOM'"
How can I solve this problem, thank you!

你好，测试集和训练集的链接都失效过期了，能否重新上传链接。

hey man.
the dataset link has been expired. please fix it.

在复现的过程中，有点难以理解有关路径设置的代码，如果您能够让 readme中的步骤更详细一些，或者出一个视频，我将非常感谢

First of all, thank you for the work you did in the paper!

I'm sorry to bother you, but I am facing some problems that I'll explain here below

Problems during Machine setup

Unfortunately, I'm encountering some troubles in running the code (and I think some versions are missing for correct reproducibility). As suggested in your README.md, I am running the experiments on ubuntu. More specifically, lsb_release -a returns:

Distributor ID:	Ubuntu
Description:	Ubuntu 22.04 LTS
Release:	22.04
Codename:	jammy

Possible error in requirements.txt or wrong cuda

The error I'm obtaining arises when i perform pip install -r requirements.txt, and it says:

Collecting torch==1.6.0 (from -r requirement.txt (line 1))
  Using cached torch-1.6.0-cp38-cp38-manylinux1_x86_64.whl (748.8 MB)
Collecting scipy==1.2.2 (from -r requirement.txt (line 2))
  Using cached scipy-1.2.2.tar.gz (23.1 MB)
  Preparing metadata (setup.py) ... done
ERROR: Ignored the following yanked versions: 0.1.6, 0.1.7, 0.1.8, 0.1.9, 0.2.0, 0.2.1, 0.2.2, 0.2.2.post2, 0.2.2.post3, 0.11.0, 0.15.0
ERROR: Ignored the following versions that require a different python version: 1.11.0 Requires-Python <3.13,>=3.9; 1.11.0rc1 Requires-Python <3.13,>=3.9; 1.11.0rc2 Requires-Python <3.13,>=3.9; 1.11.1 Requires-Python <3.13,>=3.9; 1.11.2 Requires-Python <3.13,>=3.9; 1.11.3 Requires-Python <3.13,>=3.9; 1.11.4 Requires-Python >=3.9; 1.12.0 Requires-Python >=3.9; 1.12.0rc1 Requires-Python >=3.9; 1.12.0rc2 Requires-Python >=3.9
ERROR: Could not find a version that satisfies the requirement torchvision==0.4.0 (from versions: 0.5.0, 0.6.0, 0.6.1, 0.7.0, 0.8.0, 0.8.1, 0.8.2, 0.9.0, 0.9.1, 0.10.0, 0.10.1, 0.11.1, 0.11.2, 0.11.3, 0.12.0, 0.13.0, 0.13.1, 0.14.0, 0.14.1, 0.15.1, 0.15.2, 0.16.0, 0.16.1, 0.16.2, 0.17.0)
ERROR: No matching distribution found for torchvision==0.4.0

Steps to reproduce:

On an ubuntu distribution with cuda 12.3, after having installed miniconda:

• $ git clone https://github.com/ZhouHuang23/FSPNet.git
• $ conda create -n FSPNet python=3.8
• $ conda activate FSPNet
• $ cd FSPNet/
• ~/FSPNet$ pip install -r requirement.txt

Possible missing versions

The following package / libraries versions are not specified:

• cuda
• timm
• imageio

Problems regarding results reproducibility

This problem could be due to the different versions of things I used, and I think this should be solved when the above one is solved.

Since on the server I have not enough space, I run this test on my laptop and desktop both running Arch Linux.
In order to solve the above issue, I tried to change some version of python and packages.

Through conda I installed:

• python=3.9
• cudatoolkit-dev=11.7 (adding -c conda-forge)

Through poetry I found compatible versions for python libraries, which may differ with respect to the ones you used. The pyproject.toml I declared is like the following:

[tool.poetry]
name = "fspnet"
version = "0.1.0"
description = ""
authors = ["Your Name <[email protected]>"]
readme = "README.md"

[tool.poetry.dependencies]
python = "^3.9"
torch = "1.7.1"
torchvision = "0.8.2"
scipy = "^1.12.0"
opencv-python = "4.4.0.46"
timm = "0.6.5"
imageio = "^2.34.0"


[build-system]
requires = ["poetry-core"]
build-backend = "poetry.core.masonry.api"

I tried to run predictions on COD10K test set, but seem being quite different with respect to the result you obtained. I suspect that this may be a problem of packages versions, and should be solved when the above problem is solved. Issues persist both when using a GTX 1050 (laptop configuration) and RTX 4080 (Desktop configuration)

Thank you in advance for your availability

下一步工作能否把这个工作轻量化一下，八块卡搞不起啊，普通实验室没有八块卡

作者你好, 能不能先放出数据集的预测结果图片?

非常感谢您杰出的工作！
在拜读您的代码的时候，我发现FSPNet模型的参数量非常庞大，在面对COD10K数据集仅有4040张图片的情况下，您是如何应对模型过拟合的问题的呢？
期待您的回复！

作者您好，本文的数据集链接已经失效，您是否能再上传一次数据集下载链接呢，感谢！

Thanks for the great work! I would like to ask if you can publish the code and dataset of the paper?

你们好，很高兴在你们优秀的论文中学习，整个Modle的可视化做的也非常好看。我在阅读过程中有一个问题不知道你们可不可以解决一下。
1.基于vision transfomer block的特征提取中，如下图所示，有多箭头输入NL-TEM，这一步在论文中没有提及怎么做的，具体是几个token传入呢？transfomer block的n设置为多少？这两个问题与我下面的想问的问题也有关

2.文中说FSD中总共用了12个AMI，从12-6-3-2-1。似乎初始输入给decoder的token数量为12，这样才与论文描述相符合。按照我得理解，NL-TEM对于n个输入的token可以生成n个对应的加强过locality的image。那么结论是输入给NL-TEM的tokens length 理应也是12。这很奇怪，在回答问题一transfomer block输出是什么情况以后我还想了解，NL-TEM到SFD之前是否通过了MLP或者其他手段使得维度降维到12？

希望能得到您的关注，谢谢！

您好！我是个小白，我复现过的模型不太多。这是我复现本模型的结果：

FSPNet CAMO {'Smeasure': 0.6412142300529124, 'wFmeasure': 0.4609584203873405, 'MAE': 0.10479243699011927, 'adpEm': 0.854775746954227, 'meanEm': 0.5978418087679755, 'maxEm': 0.8719062504623182, 'adpFm': 0.7269412533610914, 'meanFm': 0.5014890922791113, 'maxFm': 0.710552074958497}
FSPNet CHAMELEON {'Smeasure': 0.5321738786553841, 'wFmeasure': 0.2504145440023048, 'MAE': 0.1047958639684927, 'adpEm': 0.7769811351738184, 'meanEm': 0.4261553059051292, 'maxEm': 0.8299053763792742, 'adpFm': 0.6864579060973318, 'meanFm': 0.2954627061340793, 'maxFm': 0.7024191234422722}
FSPNet COD10K {'Smeasure': 0.6179598532498279, 'wFmeasure': 0.3396345387173911, 'MAE': 0.05923583931616759, 'adpEm': 0.8080675290969681, 'meanEm': 0.5388080955320953, 'maxEm': 0.8479816082209328, 'adpFm': 0.5808884879647634, 'meanFm': 0.36591529738642964, 'maxFm': 0.5983369130399978}

请问一下，有什么原因可能导致结果差距较大？

大佬帮忙看看