yangerkun

Results 11 comments of yangerkun

Hi, actually, the training and testing sets are split randomly, so if you have the whole dataset, you can split it by yourself, and the results should be similar.

> Would you please kindly share a copy of FLICKR-25K dataset? The link on its official site is broken. Hi, please provide your email address. I can send you a...

You can download the vgg16.npy from here : https://mega.nz/#!YU1FWJrA!O1ywiCS2IiOlUCtCpI6HTJOMrneN-Qdv3ywQP5poecM

> Hello, there is no image_train.txt file in the mirflickr dataset I downloaded. Can you give me a nuswide and mirflickr link As indicated by the Readme, image_train.txt only contains...

> But the label and path in the code seem to be divided into two txt files Yes, I check the code and they should be indicated in different files....

> So some changes was done to the code 4 days ago by the author, where he seems to have replaced cifar10 text files pointing to database, query and train...

Please give more details to find the exact problem

> 您好,为什么我用您的模型本地加载cifar-10的数据来跑的时候,跑出来的map只有0.2是正常的吗,还是说我的实验运行您的代码的时候有问题 你测试Flickr和nuswide结果如何?

那应该没问题,文中报的是map@5000的结果,可能比map@all 高一些。cifar10分辨率太低,可能结果就是不太好。

> > 那应该没问题,文中报的是map@5000的结果,可能比map@all 高一些。cifar10分辨率太低,可能结果就是不太好。 > > > 那应该没问题,文中报的是map@5000的结果,可能比map@all 高一些。cifar10分辨率太低,可能结果就是不太好。 > > > 那应该没问题,文中报的是map@5000的结果,可能比map@all 高一些。cifar10分辨率太低,可能结果就是不太好。 > > 嗯嗯,谢谢您的解答了,我查阅文献的时候看到您19年cvpr也有一篇论文distillhash, 最近在找19年对比方法,19年无监督单模态文章的方法很少,这篇论文的代码方便提供一下吗?😇 其他project的事不在这里讨论,请邮件联系哈