wendellgul / dcmh Goto Github PK

感谢你提供DCMH的PYTORCH代码，你提供了FLICKR-25K.mat的下载链接，非常感谢，请问可否提供另外两个数据集（IAPR TC-12、NUS-WIDE）的下载链接？或者合并文件的方法?
还有导入和预处理另两个数据集的代码方便提供吗？

Hello, @WendellGul thanks for the code.

I am trying to use your code, unfortunately, I am unable to download the pre-trained model from the Baidu server.

I tried using "imagenet-vgg-f.mat" fromhttp://www.vlfeat.org/matconvnet/pretrained/ but I am getting the following error:
File "/home/codes/DCMH_pytorch/models/img_module.py", line 63, in _init self.mean = torch.from_numpy(data['normalization'][0][0][0].transpose()).type(torch.float) KeyError: 'normalization'
Is there any other way you can share the pre-trained model?

Hello, do you have the processed MAT files of IAPR TC-12 and NUS-wide data sets?
Thank you very much and I am looking forward to your reply

One question: "what's the file 'file.py' imported in main.py". Thank you

您好，非常感谢您的pytorch实现。
我使用了您的代码并进行了一定的修改，使其能够通过Dataloader按需加载数据，而不是一次装载进内存。
我想请问一下您可以复现原论文的结果吗（Flickr25k数据集）？大概需要迭代多少次？
因为我无法复现论文的结果，我想知道是不是作者的问题亦或者是我自己修改时导致的问题？
在训练过程中，loss函数的term1和term2基本保持不变，仅term3能显著降低。但是最后的mAP值并没有取得效果。
logloss(term1) tensor(69321232., device='cuda:0')
quantization(term2) tensor(301857.3125, device='cuda:0')
balance(term3) tensor(924399.3750, device='cuda:0')
经过500次epoch之后，检索测试的mAP值基本与随机模型直接进行检索测试的mAP值一致，为55%，相当于没有学习到有用的信息。
我不知道应该是哪里出了问题，想请您帮我看看。最有可能出问题的地方应该在哪里？

Hi,
I also re-implement DCMH myself, and it can re-produce the performance of flickr.
But it fails to re-produce the performance of NUS-WIDE, sticking at about 0.32 and 0.34 in mAP.
can you reproduce the performance of NUS-WIDE ?