About Binomial loss hyper-parameter alpha & beta

[DJacobJiang]

How to select appropriate Alpha and Beta in Binomialloss for different dataset.Thx lot!

[DJacobJiang]

And i changed the backbone from BN-Inception to Resnet50&152 .The code is work but top R@1 just achieved 78.61 80.03 in car196.Could you please give me some Suggestions on adjusting the parameters or network structure?
Thank you so much！

[bnu-wangxun]

You can grid search the Alpha and Beta in Binomial loss for different datasets.
For small datasets as Car196, CUB200, the tricks like freeze BN, can really improve the performance a lot.
By the way, how many dimensions do you use for ResNet50&152?

[DJacobJiang]

Thank you for your reply!
In the experiment, I used 256,512,1024 as the embedding size respectively.
512 dimension seems best, but r@1 is also limited to around 80.

[bnu-wangxun]

The code is built 3 years ago. And I have open source a new coder repo for MS loss and XBM:

repo-1: https://github.com/MalongTech/research-ms-loss

repo-2: https://github.com/MalongTech/research-xbm

You can reimplement Binomial loss following the MS loss in repo-1, and the result maybe better.