The mAP of ImageNet is different from that on paper about hashnet

[caozhangjiex]

The pytorch version of the code perform badly on ImageNet. We do not find the reason yet. Please use the caffe version if you want to run on ImageNet.

[caozhangjiex]

check caffe directory

[chenshen03]

I just directly use your pretrained caffe's model to evaluate the mAP on ImageNet, and some results are as follows;

• 16bit: 0.4643
• 32bit: 0.5934

However, the results on your paper are as follows:

What makes this difference?

[caozhangjie]

Maybe it is because the caffe framework changes or other reason such as initialization. I'm sure it can reproduce the results at publishing.

[chenshen03]

I retrained the model and obtain an ideal result. Thank you~

[TreezzZ]

I retrained the model and obtain an ideal result. Thank you~

请问是如何训练的？我用PyTorch在DSH、DHN和HashNet三个算法上复现，均是在Imagenet数据集上表现奇差。

[chenshen03]

I retrained the model and obtain an ideal result. Thank you~

请问是如何训练的？我用PyTorch在DSH、DHN和HashNet三个算法上复现，均是在Imagenet数据集上表现奇差。

我之前跑过HashNet的作者提供的PyTorch版本，也自己复现过，但是在ImageNet上表现都不好，但我感觉你的表现好像比我的还更差。不过如果直接跑caffe的版本的话，结果和论文差不多。