Tensorflow implementation of the Support Vector Guided Softmax Loss for Face Recognition paper (https://arxiv.org/pdf/1812.11317.pdf).
You can see the loss implementation details and related math in this notebook. It's better to watch it on your local computer or through the nbviewer, like this than on github because of the problems with the display of the notebooks. You can also find losses here.
We use cifar-10 and cifar-100 to validate implementation and very simple cnn-model.
Cifar-100 train and test curves:
SV-AM-softmax doesn't show improvement on the cifar-10 and cifar-100 datasets. Maybe it's better works in Face Recognition problem, so it's a good idea to test it on such problems.