Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

为什么跑出来的结果与原论文给出的结果差别很大 #17

Open
adjustself opened this issue Sep 23, 2023 · 19 comments
Open

Comments

@adjustself
Copy link

adjustself commented Sep 23, 2023

1971fa65d7b90bbf8d7ed2b48545da0

@adjustself
Copy link
Author

也按照原论文代码默认的,训练了100轮,损失降到了0.01左右,为什么最后测试的效果离论文写的差距那么大

@adjustself
Copy link
Author

adjustself commented Oct 12, 2023 via email

@dehaozhou
Copy link

6acc6a7e63aadc77e4a974ed9bed54f

I have also made an attempt; however,Currently, the training is ongoing, and I am employing a strategy of using a small learning rate for multiple epochs. These results are from the sixtieth epoch. Judging by the current loss, the final outcome is not expected to surpass the current state significantly, and there is a considerable deviation from the original paper. I've observed that the primary issue lies in a specific class having a notably low Intersection over Union (IoU), which is dragging down the overall results. My approach involves splitting the entire dataset into an 80:20 ratio, but it seems that the dataset distribution might be a key factor, as the paper referenced prior experiences for dataset construction.

@adjustself
Copy link
Author

6acc6a7e63aadc77e4a974ed9bed54f I have also made an attempt; however,Currently, the training is ongoing, and I am employing a strategy of using a small learning rate for multiple epochs. These results are from the sixtieth epoch. Judging by the current loss, the final outcome is not expected to surpass the current state significantly, and there is a considerable deviation from the original paper. I've observed that the primary issue lies in a specific class having a notably low Intersection over Union (IoU), which is dragging down the overall results. My approach involves splitting the entire dataset into an 80:20 ratio, but it seems that the dataset distribution might be a key factor, as the paper referenced prior experiences for dataset construction.

这个问题我已经解决了,将epoch调到150轮左右,出来的结果基本与原论文接近

@dehaozhou
Copy link

请问下您就是用学习率0.01,batch size为8,训练了150轮得到的结果嘛?我目前batch_size为48,学习率0.005,训练了160轮,每隔20轮测试一次。发现我前天给您的回复也就是第60轮的结果是最好的,后面测试的miou值会逐渐降低,估计是过拟合了。我现在在往之前的轮次测试,看看测试结果能否提升。

@adjustself
Copy link
Author

请问下您就是用学习率0.01,batch size为8,训练了150轮得到的结果嘛?我目前batch_size为48,学习率0.005,训练了160轮,每隔20轮测试一次。发现我前天给您的回复也就是第60轮的结果是最好的,后面测试的miou值会逐渐降低,估计是过拟合了。我现在在往之前的轮次测试,看看测试结果能否提升。

我初始学习率就是0.01,然后按照源代码那样递减,batchsize就是8

@dehaozhou
Copy link

您对数据集的处理是按照原论文所述处理得嘛,还是直接把全部的混在一起然后随机分配训练集和测试集?

@dehaozhou
Copy link

大概一轮需要训练多长时间请问您还记得吗?

@adjustself
Copy link
Author

您对数据集的处理是按照原论文所述处理得嘛,还是直接把全部的混在一起然后随机分配训练集和测试集?

原论文处理的

@adjustself
Copy link
Author

vai数据集4070ti训练一次大概需要17个小时

@dehaozhou
Copy link

好的,感谢您的回复,我再进行一下尝试

@Candice-Y
Copy link

好的,感谢您的回复,我再进行一下尝试

请问SoftPool如何导入呢,能否请教一下

@dehaozhou
Copy link

dehaozhou commented Jan 27, 2024 via email

@WAutomne
Copy link

WAutomne commented Apr 7, 2024

你好,我将epoch调整到150轮左右但是最后一类car的iou特别低才0.22,你知道为什么吗??谢谢
Uploading 1712455779402.png…

@yxzz1121
Copy link

soft_cuda how to install?

@yxzz1121
Copy link

请问下您就是用学习率0.01,batch size为8,训练了150轮得到的结果嘛?我目前batch_size为48,学习率0.005,训练了160轮,每隔20轮测试一次。发现我前天给您的回复也就是第60轮的结果是最好的,后面测试的miou值会逐渐降低,估计是过拟合了。我现在在往之前的轮次测试,看看测试结果能否提升。

我初始学习率就是0.01,然后按照源代码那样递减,batchsize就是8

你是解决softpool_cuda的问题的,一直搞不好这个问题

@adjustself
Copy link
Author

请问下您就是用学习率0.01,batch size为8,训练了150轮得到的结果嘛?我目前batch_size为48,学习率0.005,训练了160轮,每隔20轮测试一次。发现我前天给您的回复也就是第60轮的结果是最好的,后面测试的miou值会逐渐降低,估计是过拟合了。我现在在往之前的轮次测试,看看测试结果能否提升。

我初始学习率就是0.01,然后按照源代码那样递减,batchsize就是8

你是解决softpool_cuda的问题的,一直搞不好这个问题

就像上面说的,win系统安装这个库非常麻烦,因为下载下来的是源代码,没有经过编译,而这个需要用到c++里面的一些依赖库编译这个源代码,我记得需要先安装VS2017,然后cuda版本对应,后面还需要配置啥环境,网上应该有相关的解答

@lgdlwy
Copy link

lgdlwy commented Jun 9, 2024

我处理的很简答,就是将原数据集从上到下按照256256的尺寸进行切割,步幅你可以自己确定,我一张原图基本上要切割几百张 不过六级不改名 @.**  

------------------ 原始邮件 ------------------ 发件人: "XinnHe/ST-UNet" @.>; 发送时间: 2023年10月12日(星期四) 下午2:58 @.>; @.@.>; 主题: Re: [XinnHe/ST-UNet] 为什么跑出来的结果与原论文给出的结果差别很大 (Issue #17) 也按照原论文代码默认的,训练了100轮,损失降到了0.01左右,为什么最后测试的效果离论文写的差距那么大 @.*** — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

你好,想请问下postdam数据集是不是标签有问题,官网说只有6个类别,实际上4-12和6-7的label有超出类别的颜色,是下载数据集有问题吗?谢谢解答

@SherryZeak
Copy link

6acc6a7e63aadc77e4a974ed9bed54f我也做了一次尝试;但是,目前,训练正在进行中,我正在采用对多个 epoch 使用小学习率的策略。这些结果来自第 60 个纪元。从目前的亏损来看,预计最终结果不会明显超过当前状态,与原始论文存在相当大的偏差。我观察到,主要问题在于特定类的交集与并集 (IoU) 非常低,这拖累了整体结果。我的方法涉及将整个数据集拆分为 80:20 的比例,但数据集分布似乎可能是一个关键因素,因为论文引用了先前的数据集构建经验。

这个问题我已经解决了,将epoch调到150轮左右,出来的结果基本与原论文接近

您好;我有几个问题想问您,数据集是怎么处理的?我试了网上关于Pot和Vai数据集的处理,但没有处理好。如果能够得到您的帮助,我会非常感激您。我们团队曾诺,从您这获得的信息不会作为学习相关的创新点,只作为对比作用。如果方便的话,可以获取您处理后的数据集或者数据集处理代码,请邮箱联系:xcj6523@163.com

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

7 participants