youngkyunjang / deep-hash-distillation Goto Github PK

Thanks for your great works!

I noticed that the testing result of mAP in NUS-WIDE dataset is different with ITQ and SH from your papar "Self-supervised Product Quantization for Deep Unsupervised Image Retrieval - ICCV2021".

the result in this paper:

the result in old paper:

I also read some other paper but they are all different. So I want to know the accurate results and reasons for this phenomenon.

Additional, I noticed that the result of this paper is much better than SPQ. Does it means unsupervised pq-based methods still has much space to improve?

Hi, thank you for awesome project.
I've tried your experiment setting, Max map produced by the model is lower than the value in the paper,
eg. ImageNet AlexNet bit 64 is 0.716 vs 0.721

I tried several times and still cannot get higher map, do you have any other specific hyper-parameter settings?

When I try to reproduce the result using SwinTransformer, I got this error

AttributeError: 'SwinTransformer' object has no attribute 'avgpool'

the problem is from Swintransformer doesn't have avgpool layer

code that call avgpool

Is there a different pre-trained model used in this project or can I remove avgpool layer?

Hi, I'm a fresh learner of hash, and just cloned your repo to my server.
Under the imagent-resnet setting, the code can not work.
It seems like need to resize imagenet's images to universal resolution first. Cause the images in orginal imagenet database has various resolution. And pytorch raise a exception when use your custom dataloader.
stack expects each tensor to be equal size, but got [3, 375, 500] at entry 0 and [3, 427, 640] at entry 1
By the way, your work is awesome!

Awesome work, and thank you for releasing your code publicly. I would like to know if we can get access to pre-trained weights of Swin-T and Resent for Imagenet and MSCOCO. That will be very helpful for our work and save us lots of training time.

I observed that your repository does not seem to provide scripts for multi-gpu training. May I ask if your current code can only train on a single gpu?

Great job. I have a question, how to output the top N most similar images based on the query image.

Hello,
Recently, I am deeply studying about image retrieval, and I want to exercise my ability through this code. I read it carefully and downloaded coco2014 according to the requirements of readme.md, but when I run train.py, many of the labels in. /data/txt are different from the image file names in datasets, showing that there is no file. I want to know how the code in labels is formed and can it be changed? In addition, I don't know which year to download ImageNet and NUS-WIDE datasets. I'm confused and don't know what to do. I hope I can get your help, thank you very much!