kamwoh / orthohash Goto Github PK

Thank you very much for your work. I am very interested in showing the Web interface in the video. Is there any code for this part?

In face recognition, when the number of classes exceeds 10 million and K=128, the CosFace/ArcFace loss function hardly converges, due to the fixed binary orthogonal codes.
Any suggestions to help with training?
Thanks a lot!

Hi~ Thanks for your excellent work. I have a question about

Thanks for sharing the code, but When I followed the instruction and got ready to test it on COCO I met some error

RuntimeError: stack expects each tensor to be equal size, but got [3, 297, 224] at entry 0 and [3, 224, 335] at entry 1

At first, I think the question was about the transform

'coco': [
                    transforms.Resize(resize),
                    transforms.RandomCrop(crop),
                    transforms.RandomHorizontalFlip()
                ]

but I tried several times to change the transform, still wrong, would you please do me a favor?

Hi~ Thanks for sharing the source code.

I noticed you sample from a Bernoulli distribution to generate the codebook but seems the codebook is not updated along with the training.
Would it be better to update it? Or is it unnecessary to do so? Do you have any theoretical proof on this part?

And also I noticed you emphasize orthogonal several times in your paper, does that mean as long as the orthogonality is ensured the updating of the codebook is not needed?