Embedding-Network
Embedding-Network copied to clipboard
On the latest version (VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically
To adapt to the new version of VehicleID dataset, I change a few lines of the code in train_soft_trip_v2.py:
and comment the following two lines:
in VehicleID.py
Did I do something wrong somewhere? or this is the true results on the new dataset? Thanks.
你好,请问这个问题你解决了吗,我是修改了数据集,修改后用记事本和写字板打开分别是这样的,可以运行,但训练出的模型评价超级差,MAP大概是0.05,想知道你是怎么解决的呢~~谢谢!
昨天才开始跑,还没有解决。你这个这么低,估计数据读入有问题
好的谢谢奥!我再看看~
------------------ 原始邮件 ------------------ 发件人: "Andybert"[email protected]; 发送时间: 2019年7月5日(星期五) 上午10:42 收件人: "yanbai1993/Embedding-Network"[email protected]; 抄送: "Cherish"[email protected];"Comment"[email protected]; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
昨天才开始跑,还没有解决。你这个这么低,估计数据读入有问题
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
请注意数据集的格式,因为使用了softmax,所有车辆id 必须从0开始编写,并注意分类器的类别数量和id 总数量是否一致。正常的训练和测试流程,在最新的vehicleID_v1.0 上,800的规模,可以获得82%的mAP
谢谢!找到问题了,训练测试都没有问题,只是因为我用只训练了10000次时保存的模型进行测试的,换成训练50000次的模型,rank-1就达到65.8%
谢谢!!我再看一看!!
您好!我又来打扰了。您代码只改了上图两三个地方吗。我原来是先改的数据集又训练的,精度很差,我用了您上面的方法,直接在代码中修正数据集精度还是很差,训练过程是很正常的,没有出错什么的,所以请教您有没有遇见这种问题的解决经验呢。
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
嗯嗯是的,是VehicleID_V1.0,您记得训练的时候它的loss大概是从多少开始收敛的嘛,我在这里triplet大概是从1,softmax大概是从9~
------------------ 原始邮件 ------------------ 发件人: "Andybert"[email protected]; 发送时间: 2019年7月10日(星期三) 中午12:45 收件人: "yanbai1993/Embedding-Network"[email protected]; 抄送: "陈大猫子"[email protected];"Comment"[email protected]; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
要用imagenet的模型初始化
发自我的 iPhone
在 2019年7月10日,15:03,chenchenver <[email protected]mailto:[email protected]> 写道:
嗯嗯是的,是VehicleID_V1.0,您记得训练的时候它的loss大概是从多少开始收敛的嘛,我在这里triplet大概是从1,softmax大概是从9~
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 中午12:45 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBQUV67LQICVHG5UPTDP6WCS7A5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSQIHI#issuecomment-509936669, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBV23XRCWPOYYSIPPBTP6WCS7ANCNFSM4H6DZMLQ.
所以不能直接按照readme上面的步骤直接处理vehicleID数据集是嘛
------------------ 原始邮件 ------------------ 发件人: "Andybert"[email protected]; 发送时间: 2019年7月10日(星期三) 下午3:19 收件人: "yanbai1993/Embedding-Network"[email protected]; 抄送: "陈大猫子"[email protected];"Comment"[email protected]; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
要用imagenet的模型初始化
发自我的 iPhone
在 2019年7月10日,15:03,chenchenver <[email protected]mailto:[email protected]> 写道:
嗯嗯是的,是VehicleID_V1.0,您记得训练的时候它的loss大概是从多少开始收敛的嘛,我在这里triplet大概是从1,softmax大概是从9~
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 中午12:45 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBQUV67LQICVHG5UPTDP6WCS7A5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSQIHI#issuecomment-509936669, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBV23XRCWPOYYSIPPBTP6WCS7ANCNFSM4H6DZMLQ.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
数据集正常处理,确认一下你们训练的时候有没有用imagenet初始化resnet50的网络。这个对性能影响较大
发自我的 iPhone
在 2019年7月10日,15:27,chenchenver <[email protected]mailto:[email protected]> 写道:
所以不能直接按照readme上面的步骤直接处理vehicleID数据集是嘛
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 下午3:19 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
要用imagenet的模型初始化
发自我的 iPhone
在 2019年7月10日,15:03,chenchenver <[email protected]mailto:[email protected]mailto:[email protected]> 写道:
嗯嗯是的,是VehicleID_V1.0,您记得训练的时候它的loss大概是从多少开始收敛的嘛,我在这里triplet大概是从1,softmax大概是从9~
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 中午12:45 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBQUV67LQICVHG5UPTDP6WCS7A5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSQIHI#issuecomment-509936669, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBV23XRCWPOYYSIPPBTP6WCS7ANCNFSM4H6DZMLQ.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBSHF2ZCANGPMJJ4LITP6WFPNA5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSSBHQ#issuecomment-509943966, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBQMDH7T75LX3OAJDGLP6WFPNANCNFSM4H6DZMLQ.
谢谢奥!训练的过程中有从param_url = 'https://download.pytorch.org/models/resnet50-19c8e357.pth'下载预训练模型,其他的我再仔细看下!!真的非常谢谢你的回复!
------------------ 原始邮件 ------------------ 发件人: "Andybert"[email protected]; 发送时间: 2019年7月10日(星期三) 下午3:29 收件人: "yanbai1993/Embedding-Network"[email protected]; 抄送: "陈大猫子"[email protected];"Comment"[email protected]; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
数据集正常处理,确认一下你们训练的时候有没有用imagenet初始化resnet50的网络。这个对性能影响较大
发自我的 iPhone
在 2019年7月10日,15:27,chenchenver <[email protected]mailto:[email protected]> 写道:
所以不能直接按照readme上面的步骤直接处理vehicleID数据集是嘛
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 下午3:19 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
要用imagenet的模型初始化
发自我的 iPhone
在 2019年7月10日,15:03,chenchenver <[email protected]mailto:[email protected]mailto:[email protected]> 写道:
嗯嗯是的,是VehicleID_V1.0,您记得训练的时候它的loss大概是从多少开始收敛的嘛,我在这里triplet大概是从1,softmax大概是从9~
------------------ 原始邮件 ------------------ 发件人: "Andybert"<[email protected]mailto:[email protected]mailto:[email protected]>; 发送时间: 2019年7月10日(星期三) 中午12:45 收件人: "yanbai1993/Embedding-Network"<[email protected]mailto:[email protected]mailto:[email protected]>; 抄送: "陈大猫子"<[email protected]mailto:[email protected]mailto:[email protected]>;"Comment"<[email protected]mailto:[email protected]mailto:[email protected]>; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
代码确实只修改了那几个地方,没有修改数据集(直接用新版数据集里面提供的训练集和测试集列表文件),其他都是按照说明文档来操作的。你也是在VehicleID_V1.0上训练和测试的?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBQUV67LQICVHG5UPTDP6WCS7A5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSQIHI#issuecomment-509936669, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBV23XRCWPOYYSIPPBTP6WCS7ANCNFSM4H6DZMLQ.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yanbai1993/Embedding-Network/issues/5?email_source=notifications&email_token=ADZPKBSHF2ZCANGPMJJ4LITP6WFPNA5CNFSM4H6DZML2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZSSBHQ#issuecomment-509943966, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADZPKBQMDH7T75LX3OAJDGLP6WFPNANCNFSM4H6DZMLQ.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
我的soft_loss太大了,500iter的时候还在9左右,现在我重新训练了下10000多了也才到2
------------------ 原始邮件 ------------------ 发件人: "Andybert"[email protected]; 发送时间: 2019年7月10日(星期三) 下午3:43 收件人: "yanbai1993/Embedding-Network"[email protected]; 抄送: "陈大猫子"[email protected];"Comment"[email protected]; 主题: Re: [yanbai1993/Embedding-Network] On the latest version(VehicleID_V1.0) of VehicleID, the accuracy dropped dramatically (#5)
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.