-
Notifications
You must be signed in to change notification settings - Fork 233
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
请问大家seg_net训练了多久? #53
Comments
segnet在BS=16 GPU:P100-12GB需要5个半小时 |
好的,谢谢,有可能是我的硬件问题,运行时间比较长,那么请问你的segnet和unet的代码是不是一样
…------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
请问预测结果是这样的,是什么原因呐?非常感谢
…------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
…________________________________
From: Bodhi-robbery <[email protected]>
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A>.
获取 Outlook for Android<https://aka.ms/ghei36>
|
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
…------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年11月24日(星期天) 下午2:34
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A>.
获取 Outlook for Android<https://aka.ms/ghei36>
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36>
…________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年11月24日(星期天) 下午2:34
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A>.
获取 Outlook for Android<https://aka.ms/ghei36>
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A>.
|
那请问是不是unet不适合用于多分类?或者说效果根本不如二分类?
…------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年11月24日(星期天) 下午2:57
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36>
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------&nbsp;原始邮件&nbsp;------------------
发件人:&nbsp;"Shikairan"<[email protected]&gt;;
发送时间:&nbsp;2019年11月24日(星期天) 下午2:34
收件人:&nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送:&nbsp;"-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题:&nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]&gt;
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]&gt;
Cc: Shikairan <[email protected]&gt;; Comment <[email protected]&gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]&gt;;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送: "-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A&gt;.
获取 Outlook for Android<https://aka.ms/ghei36&gt;
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A>.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
模型的多分类十分依赖于labe的标注精度,一个像素点在多分类模型下分类错误的可能性更高,在被分类错误的情况下还虑人为考虑该像素点应该是属于哪个类别;而二分类只需要考虑“是”与“不是”的问题,是非黑即白的处理方式。麻烦的地方在于每个种类都需要一个模型,但不在依赖于高精度的label标注。多分类可以一个模型解决所有分类,但是依赖于高精度label。
获取 Outlook for Android<https://aka.ms/ghei36>
…________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 3:21:59 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
那请问是不是unet不适合用于多分类?或者说效果根本不如二分类?
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]>;
发送时间: 2019年11月24日(星期天) 下午2:57
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "->慧慧<-"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36>
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------&nbsp;原始邮件&nbsp;------------------
发件人:&nbsp;"Shikairan"<[email protected]&gt;;
发送时间:&nbsp;2019年11月24日(星期天) 下午2:34
收件人:&nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送:&nbsp;"-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题:&nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]&gt;
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]&gt;
Cc: Shikairan <[email protected]&gt;; Comment <[email protected]&gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]&gt;;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送: "-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A&gt;.
获取 Outlook for Android<https://aka.ms/ghei36&gt;
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A>.
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3BS4QGIVJSGU2A4ZS3QVITRPA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFLNI#issuecomment-557864373>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DHA6KP6DZ46XSTDWDQVITRPANCNFSM4I3SB56A>.
|
所以说unet不是很适用于多分类,谢谢你的回复。请问您知不知道常用于二分类的数据集。非常感谢!
…---原始邮件---
发件人: "Shikairan"<[email protected]>
发送时间: 2019年11月24日(星期日) 下午3:36
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "Bodhi-robbery"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
模型的多分类十分依赖于labe的标注精度,一个像素点在多分类模型下分类错误的可能性更高,在被分类错误的情况下还虑人为考虑该像素点应该是属于哪个类别;而二分类只需要考虑“是”与“不是”的问题,是非黑即白的处理方式。麻烦的地方在于每个种类都需要一个模型,但不在依赖于高精度的label标注。多分类可以一个模型解决所有分类,但是依赖于高精度label。
获取 Outlook for Android<https://aka.ms/ghei36>
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 3:21:59 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
那请问是不是unet不适合用于多分类?或者说效果根本不如二分类?
------------------ 原始邮件 ------------------
发件人:&nbsp;"Shikairan"<[email protected]&gt;;
发送时间:&nbsp;2019年11月24日(星期天) 下午2:57
收件人:&nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送:&nbsp;"-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题:&nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36&gt;
________________________________
From: Bodhi-robbery <[email protected]&gt;
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]&gt;
Cc: Shikairan <[email protected]&gt;; Comment <[email protected]&gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人:&amp;nbsp;"Shikairan"<[email protected]&amp;gt;;
发送时间:&amp;nbsp;2019年11月24日(星期天) 下午2:34
收件人:&amp;nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&amp;gt;;
抄送:&amp;nbsp;"-&amp;gt;慧慧<-"<[email protected]&amp;gt;;"Author"<[email protected]&amp;gt;;
主题:&amp;nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]&amp;gt;
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]&amp;gt;
Cc: Shikairan <[email protected]&amp;gt;; Comment <[email protected]&amp;gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]&amp;gt;;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&amp;gt;;
抄送: "-&amp;gt;慧慧<-"<[email protected]&amp;gt;;"Author"<[email protected]&amp;gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;amp;email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060&amp;gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A&amp;gt;.
获取 Outlook for Android<https://aka.ms/ghei36&amp;gt;
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A&gt;.
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3BS4QGIVJSGU2A4ZS3QVITRPA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFLNI#issuecomment-557864373>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DHA6KP6DZ46XSTDWDQVITRPANCNFSM4I3SB56A>.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
抱歉,数据是保密资料,不能共享。
获取 Outlook for Android<https://aka.ms/ghei36>
…________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 3:54:32 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
所以说unet不是很适用于多分类,谢谢你的回复。请问您知不知道常用于二分类的数据集。非常感谢!
---原始邮件---
发件人: "Shikairan"<[email protected]>
发送时间: 2019年11月24日(星期日) 下午3:36
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "Bodhi-robbery"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
模型的多分类十分依赖于labe的标注精度,一个像素点在多分类模型下分类错误的可能性更高,在被分类错误的情况下还虑人为考虑该像素点应该是属于哪个类别;而二分类只需要考虑“是”与“不是”的问题,是非黑即白的处理方式。麻烦的地方在于每个种类都需要一个模型,但不在依赖于高精度的label标注。多分类可以一个模型解决所有分类,但是依赖于高精度label。
获取 Outlook for Android<https://aka.ms/ghei36>
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 3:21:59 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
那请问是不是unet不适合用于多分类?或者说效果根本不如二分类?
------------------ 原始邮件 ------------------
发件人:&nbsp;"Shikairan"<[email protected]&gt;;
发送时间:&nbsp;2019年11月24日(星期天) 下午2:57
收件人:&nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送:&nbsp;"-&gt;慧慧<-"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题:&nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36&gt;
________________________________
From: Bodhi-robbery <[email protected]&gt;
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]&gt;
Cc: Shikairan <[email protected]&gt;; Comment <[email protected]&gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人:&amp;nbsp;"Shikairan"<[email protected]&amp;gt;;
发送时间:&amp;nbsp;2019年11月24日(星期天) 下午2:34
收件人:&amp;nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&amp;gt;;
抄送:&amp;nbsp;"-&amp;gt;慧慧<-"<[email protected]&amp;gt;;"Author"<[email protected]&amp;gt;;
主题:&amp;nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]&amp;gt;
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]&amp;gt;
Cc: Shikairan <[email protected]&amp;gt;; Comment <[email protected]&amp;gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]&amp;gt;;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&amp;gt;;
抄送: "-&amp;gt;慧慧<-"<[email protected]&amp;gt;;"Author"<[email protected]&amp;gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;amp;email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060&amp;gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A&amp;gt;.
获取 Outlook for Android<https://aka.ms/ghei36&amp;gt;
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A&gt;.
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3BS4QGIVJSGU2A4ZS3QVITRPA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFLNI#issuecomment-557864373>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DHA6KP6DZ46XSTDWDQVITRPANCNFSM4I3SB56A>.
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3D72REYFUTYOQ22EALQVIXLRA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFYMQ#issuecomment-557866034>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3GWJBMPKC36Q4TUYCTQVIXLRANCNFSM4I3SB56A>.
|
非常感谢您!
…---原始邮件---
发件人: "Shikairan"<[email protected]>
发送时间: 2019年11月24日(星期日) 下午3:56
收件人: "AstarLight/Satellite-Segmentation"<[email protected]>;
抄送: "Bodhi-robbery"<[email protected]>;"Author"<[email protected]>;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,数据是保密资料,不能共享。
获取 Outlook for Android<https://aka.ms/ghei36>
________________________________
From: Bodhi-robbery <[email protected]>
Sent: Sunday, November 24, 2019 3:54:32 PM
To: AstarLight/Satellite-Segmentation <[email protected]>
Cc: Shikairan <[email protected]>; Comment <[email protected]>
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
所以说unet不是很适用于多分类,谢谢你的回复。请问您知不知道常用于二分类的数据集。非常感谢!
---原始邮件---
发件人: "Shikairan"<[email protected]&gt;
发送时间: 2019年11月24日(星期日) 下午3:36
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&gt;;
抄送: "Bodhi-robbery"<[email protected]&gt;;"Author"<[email protected]&gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
模型的多分类十分依赖于labe的标注精度,一个像素点在多分类模型下分类错误的可能性更高,在被分类错误的情况下还虑人为考虑该像素点应该是属于哪个类别;而二分类只需要考虑“是”与“不是”的问题,是非黑即白的处理方式。麻烦的地方在于每个种类都需要一个模型,但不在依赖于高精度的label标注。多分类可以一个模型解决所有分类,但是依赖于高精度label。
获取 Outlook for Android<https://aka.ms/ghei36&gt;
________________________________
From: Bodhi-robbery <[email protected]&gt;
Sent: Sunday, November 24, 2019 3:21:59 PM
To: AstarLight/Satellite-Segmentation <[email protected]&gt;
Cc: Shikairan <[email protected]&gt;; Comment <[email protected]&gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
那请问是不是unet不适合用于多分类?或者说效果根本不如二分类?
------------------ 原始邮件 ------------------
发件人:&amp;nbsp;"Shikairan"<[email protected]&amp;gt;;
发送时间:&amp;nbsp;2019年11月24日(星期天) 下午2:57
收件人:&amp;nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&amp;gt;;
抄送:&amp;nbsp;"-&amp;gt;慧慧<-"<[email protected]&amp;gt;;"Author"<[email protected]&amp;gt;;
主题:&amp;nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
我没有找到较好的传统图像处理方式,现在有两种解决方法:
1、cGAN做图像补齐
2、再做一个额外的unet对识别完的图片再做一次pix2pix
获取 Outlook for Android<https://aka.ms/ghei36&amp;gt;
________________________________
From: Bodhi-robbery <[email protected]&amp;gt;
Sent: Sunday, November 24, 2019 2:39:56 PM
To: AstarLight/Satellite-Segmentation <[email protected]&amp;gt;
Cc: Shikairan <[email protected]&amp;gt;; Comment <[email protected]&amp;gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
你好,感觉github上的代码不完全,我在进行多分类的时候训练出来的unet拼接痕迹很是明显。请问有什么解决方法吗?非常感谢你的回复!
------------------&amp;amp;nbsp;原始邮件&amp;amp;nbsp;------------------
发件人:&amp;amp;nbsp;"Shikairan"<[email protected]&amp;amp;gt;;
发送时间:&amp;amp;nbsp;2019年11月24日(星期天) 下午2:34
收件人:&amp;amp;nbsp;"AstarLight/Satellite-Segmentation"<[email protected]&amp;amp;gt;;
抄送:&amp;amp;nbsp;"-&amp;amp;gt;慧慧<-"<[email protected]&amp;amp;gt;;"Author"<[email protected]&amp;amp;gt;;
主题:&amp;amp;nbsp;Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
抱歉,最近因为保密项目呆在深山里才被放出来。segnet的代码不一样,因为Keras不能做pooling index,所以segnet代码有一部分是用tf重写的。
首先,这个git里的数据集是有问题的,标注label十分不准确,所以不必期望该数据集能训练出优秀模型。其次,使用一个模型下的多分类识别率相较多个模型的二分类的总和在精度下是相差巨大,建议做成做模型二分类。
因为segnet的消耗内存巨大,在更换数据集后,12GB的P100下BS最多只能设置成17(总训练量50000份256*256彩图数据),建筑识别率在90%,道路识别率在97%。UNET与segnet的正确率差别不太大,综合来看数据集标注影响准确率巨大。
________________________________
From: Bodhi-robbery <[email protected]&amp;amp;gt;
Sent: Saturday, October 12, 2019 9:26:26 AM
To: AstarLight/Satellite-Segmentation <[email protected]&amp;amp;gt;
Cc: Shikairan <[email protected]&amp;amp;gt;; Comment <[email protected]&amp;amp;gt;
Subject: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
请问预测结果是这样的,是什么原因呐?非常感谢
------------------ 原始邮件 ------------------
发件人: "Shikairan"<[email protected]&amp;amp;gt;;
发送时间: 2019年10月11日(星期五) 下午5:01
收件人: "AstarLight/Satellite-Segmentation"<[email protected]&amp;amp;gt;;
抄送: "-&amp;amp;gt;慧慧<-"<[email protected]&amp;amp;gt;;"Author"<[email protected]&amp;amp;gt;;
主题: Re: [AstarLight/Satellite-Segmentation] 请问大家seg_net训练了多久? (#53)
segnet在BS=16 GPU:P100-12GB需要5个半小时
Unet在BS=64 在50000个训练样本下 GPU:P100-12GB需要5个小时
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;amp;amp;email_token=AM6VO3D4SEKX43Q7PZQVFRDQOERUFA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBBRYRA#issuecomment-541269060&amp;amp;gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DBRC3PMZXGG5C64KDQOERUFANCNFSM4I3SB56A&amp;amp;gt;.
获取 Outlook for Android<https://aka.ms/ghei36&amp;amp;gt;
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;amp;email_token=AM6VO3DDM3PHLTHCH4Q32VDQVIOTZA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAE35Q#issuecomment-557862390&amp;gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3F254EJSXUEUZL6CIDQVIOTZANCNFSM4I3SB56A&amp;gt;.
D
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
D
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&amp;email_token=AM6VO3BS4QGIVJSGU2A4ZS3QVITRPA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFLNI#issuecomment-557864373&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3DHA6KP6DZ46XSTDWDQVITRPANCNFSM4I3SB56A&gt;.
―
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#53?email_source=notifications&email_token=AM6VO3D72REYFUTYOQ22EALQVIXLRA5CNFSM4I3SB56KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFAFYMQ#issuecomment-557866034>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AM6VO3GWJBMPKC36Q4TUYCTQVIXLRANCNFSM4I3SB56A>.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
|
您好,请问您在该数据集上最好的效果,评价指标做到多少哈? |
这个git里面的数据集是有问题的,用这个数据集做出来的模型名义上有70%的准确率,但是图像识别完的结果惨不忍睹,详细原因我在这个git里单独说一下。 |
是,我也发现了标注不准确问题,而且类别不均衡,真的是难办。话说您对类别不均衡有什么好的解决方法吗?(我试过 focal loss 但好像作用不明显) |
将所有label抽出来单独训练成数个二分模型,5个label就训练成5个二分模型 |
No description provided.
The text was updated successfully, but these errors were encountered: