imagecbj / A-serial-image-copy-move-forgery-localization-scheme-with-source-target-distinguishment

28 stars 8 forks source link

复现Busternet请教 #3

Open adamas-v opened 3 years ago

adamas-v commented 3 years ago

作者你好,我在复现Busternet的结果时遇到了困难,我按照论文里面的指示设定学习率(1.两个分支首先使用1e-2,其中simi-det加载vgg模型。2.当val_loss达到平台期20代时减半学习率。)可是我现在在这两个分支上已经训练了很多代,Va_simi_loss下降到了0.13,val_mani_loss下降到了0.036,目前网络在验证集上的效果很不错,但是在测试集CASIA上的效果还是很差。 不知道这个复现过程还有多久。或者有什么细节我没注意到。恳请指点。

imagecbj commented 3 years ago

你好!

你这个现象是过拟合了,具体很难指导,需要自己在网络上查找一些方法处理。 有点差异的是,我们训练时并没有加载预训练vgg模型,Busternet原文中好像也没有。

2021-05-30

陈北京 Dr. Beijing Chen School of Computer & Software, Nanjing University of Information Science & Technology, Jiangsu Province, P.R.C @.*** https://faculty.nuist.edu.cn/chenbeijing/zh_CN/index/65713/list/index.htm

发件人:Adamas @.> 发送时间:2021-05-29 16:06 主题:[imagecbj/A-serial-image-copy-move-forgery-localization-scheme-with-source-target-distinguishment] 复现Busternet请教 (#3) @.> @.***>

作者你好,我在复现Busternet的结果时遇到了困难,我按照论文里面的指示设定学习率(1.两个分支首先使用1e-2,其中simi-det加载vgg模型。2.当val_loss达到平台期20代时减半学习率。)可是我现在在这两个分支上已经训练了很多代,Va_simi_loss下降到了0.13,val_mani_loss下降到了0.036,目前网络在验证集上的效果很不错,但是在测试集CASIA上的效果还是很差。 不知道这个复现过程还有多久。或者有什么细节我没注意到。恳请指点。 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

adamas-v commented 3 years ago

非常非常感谢您的回复,我还有一个很关键的疑惑,就是用于在测试集casia上预测的模型,是否需要在casia上finetune呢,因为我有留意到casia数据集中copymove的sample是有3295个,而论文中的测试集是1313个。 还是说只在uscisi上训练就可以达到论文中在casia的测试结果呢?

imagecbj commented 3 years ago

不进行微调,只在BusterNet提供的新合成数据集上训练。

2021-06-03

陈北京 Dr. Beijing Chen School of Computer & Software, Nanjing University of Information Science & Technology, Jiangsu Province, P.R.C @.*** https://faculty.nuist.edu.cn/chenbeijing/zh_CN/index/65713/list/index.htm

发件人:Adamas @.> 发送时间:2021-05-30 23:46 主题:Re: [imagecbj/A-serial-image-copy-move-forgery-localization-scheme-with-source-target-distinguishment] 复现Busternet请教 (#3) @.> @.**@.>

非常非常感谢您的回复,我还有一个很关键的疑惑,就是用于在测试集casia上预测的模型,是否需要在casia上funetune呢,因为我有留意到casia数据集中copymove的sample是有3295个,而论文中的测试集是1313个。 还是说只在uscisi上训练就可以达到论文中在casia的测试结果呢? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

170744039 commented 2 years ago

作者你好,我在复现Busternet的结果时遇到了困难,我按照论文里面的指示设定学习率(1.两个分支首先使用1e-2,其中simi-det加载vgg模型。2.当val_loss达到平台期20代时减半学习率。)可是我现在在这两个分支上已经训练了很多代,Va_simi_loss下降到了0.13,val_mani_loss下降到了0.036,目前网络在验证集上的效果很不错,但是在测试集CASIA上的效果还是很差。 不知道这个复现过程还有多久。或者有什么细节我没注意到。恳请指点。

你好,可以交流一下吗? 我的qq2978064409