softmax vs cosine softmax ?

[YoungjaeDev]

Hi
In the official code (https://github.com/nwojke/deep_sort), he focused on generative learning using cosine_softmax loss

However, in the code implemented by pytorch that code refer to, it simply conducted a discrimitive learning with cross_entropy

deep_sort_pytorch/deep_sort/deep/train.py

Line 67 in 8cfe246

criterion = torch.nn.CrossEntropyLoss()

Is there a big difference between the reason and the performance of this implementation?

[yesmineht]

hello @youngjae-avikus did you change it to cosine softmax_loss ?