CommissarMa / Context-Aware_Crowd_Counting-pytorch

The implementation of Context-Aware Crowd Counting(CVPR2019)
MIT License
67 stars 23 forks source link

validation的选择 #8

Open visionxyz opened 5 years ago

visionxyz commented 5 years ago

请问您这个是不是只有train test,没有val?

CommissarMa commented 5 years ago

是的

---原始邮件--- 发件人: "blessxu"notifications@github.com 发送时间: 2019年7月6日(星期六) 晚上9:07 收件人: "CommissarMa/Context-Aware_Crowd_Counting-pytorch"Context-Aware_Crowd_Counting-pytorch@noreply.github.com; 抄送: "Subscribed"subscribed@noreply.github.com; 主题: [CommissarMa/Context-Aware_Crowd_Counting-pytorch] validation的选择 (#8)

请问您这个是不是只有train test,没有val?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

visionxyz commented 5 years ago

作者还在自己主页上说it successfully reproduce the performance,感觉只分train test这样不太严谨。如果我只分train和test,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了val的。可是分了val之后结果其实蛮受到这个val的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分val。

CommissarMa commented 5 years ago

从去年的csrnet开始就不分val了

---原始邮件--- 发件人: "blessxu"notifications@github.com 发送时间: 2019年7月6日(星期六) 晚上9:12 收件人: "CommissarMa/Context-Aware_Crowd_Counting-pytorch"Context-Aware_Crowd_Counting-pytorch@noreply.github.com; 抄送: "Comment"comment@noreply.github.com;"Ma Zhenwei"960901625@qq.com; 主题: Re: [CommissarMa/Context-Aware_Crowd_Counting-pytorch] validation的选择 (#8)

作者还在自己主页上说it successfully reproduce the performance,感觉只分train test这样不太严谨。如果我只分train和test,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了val的。可是分了val之后结果其实蛮受到这个val的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分val。

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

visionxyz commented 5 years ago

原来如此嘛,那我也不分了。昨天问过C-3 framework的作者。https://github.com/gjy3035/C-3-Framework/issues/35 ,在这个issue,他还是说要分val,但是看他复现的结果,很多都不如paper了。感谢您的回复。

CommissarMa commented 5 years ago

没事,只有mcnn那个是分了val的,共同进步

---原始邮件--- 发件人: "blessxu"notifications@github.com 发送时间: 2019年7月6日(星期六) 晚上9:15 收件人: "CommissarMa/Context-Aware_Crowd_Counting-pytorch"Context-Aware_Crowd_Counting-pytorch@noreply.github.com; 抄送: "Comment"comment@noreply.github.com;"Ma Zhenwei"960901625@qq.com; 主题: Re: [CommissarMa/Context-Aware_Crowd_Counting-pytorch] validation的选择 (#8)

原来如此嘛,那我也不分了。昨天问过C-3 framework的作者。gjy3035/C-3-Framework#35 ,在这个issue,他还是说要分val,但是看他复现的结果,很多都不如paper了。感谢您的回复。

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

visionxyz commented 5 years ago

哈哈哈,MCNN的作者我认识。我也问过他这个,他也说要分(难怪了)。

CommissarMa commented 5 years ago

但目前你如果想超过sota,还是别分了,因为他们也没分,祝你成功

---原始邮件--- 发件人: "blessxu"notifications@github.com 发送时间: 2019年7月6日(星期六) 晚上9:18 收件人: "CommissarMa/Context-Aware_Crowd_Counting-pytorch"Context-Aware_Crowd_Counting-pytorch@noreply.github.com; 抄送: "Comment"comment@noreply.github.com;"Ma Zhenwei"960901625@qq.com; 主题: Re: [CommissarMa/Context-Aware_Crowd_Counting-pytorch] validation的选择 (#8)

哈哈哈,MCNN的作者我认识。我也问过他这个,他也说要分(难怪了)。

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

aain3314 commented 4 years ago

作者还在自己主页上说它成功地再现了表现,感觉只分火车测试这样不太严谨。如果我只分火车和测试,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了VAL的。可是分了VAL之后结果其实蛮受到这个VAL的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分VAL。

请问csrnet网络不分val 是通过删除val.json文件实现的吗?

aain3314 commented 4 years ago

作者还在自己主页上说它成功地重现了表现,感觉只分火车测试这样不太严谨。如果我只分火车和测试,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了VAL的。可是分了VAL之后结果其实蛮受到这个VAL的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分VAL。

在csrnet中你是怎么没用val的?想请教一下,864770486@qq.com,谢谢

visionxyz commented 4 years ago

作者还在自己主页上说它成功地再现了表现,感觉只分火车测试这样不太严谨。如果我只分火车和测试,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了VAL的。可是分了VAL之后结果其实蛮受到这个VAL的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分VAL。

请问csrnet网络不分val 是通过删除val.json文件实现的吗?

那个不就不用就好了?本来就只有train test

CommissarMa commented 4 years ago

作者还在自己主页上说它成功地重现了表现,感觉只分火车测试这样不太严谨。如果我只分火车和测试,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了VAL的。可是分了VAL之后结果其实蛮受到这个VAL的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分VAL。

在csrnet中你是怎么没用val的?想请教一下,864770486@qq.com,谢谢

你就把本来用val.json的地方全部换成test.json就行了。

eaglex-sh commented 4 years ago

作者还在自己主页上说it successfully reproduce the performance,感觉只分train test这样不太严谨。如果我只分train和test,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了val的。可是分了val之后结果其实蛮受到这个val的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分val。

把CSRNet训到61.3是end2end的吗?是不是直接把val换成test,run就行了?用的是pytorch版本的吗?我用的pytorch版本好像训练不到那么低,也就66+吧感觉

visionxyz commented 4 years ago

我有模型外的改進,CSR还是CSR

发自我的iPhone

在 2019年10月16日,13:36,guodewen notifications@github.com 写道:

 作者还在自己主页上说it successfully reproduce the performance,感觉只分train test这样不太严谨。如果我只分train和test,我现在能把CSRNet训练到61.3,这让我很疑惑。感觉这些论文还是应该分了val的。可是分了val之后结果其实蛮受到这个val的影响的。我问了好几个论文的作者,还没有人回复我他们到底分没分val。

把CSRNet训到61.3是end2end的吗?是不是直接把val换成test,run就行了?用的是pytorch版本的吗?我用的pytorch版本好像训练不到那么低,也就66+吧感觉

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub, or unsubscribe.