error nan loss

[damvantai]

when i use arcface loss in mnist, my error loss nan ? do you test your code accuracy ?

[auroua]

I have not test the code on mnist. I think you can try the following method to make arcface work probably on mnist.

1. Try to use small network.
2. The 50-layer network use lots of batch norm layer, but the mnist dataset only contain 10 category and the variance in each category is not large enough to fit so many batch norm layers, so it will cause the nan. You should to reduce the number of batch norm layers.

[damvantai]

i am using small neural network, but loss error? nan https://github.com/damvantai/tensorflow-sphereface-asoftmax/blob/master/test_mnist.py

[damvantai]

i use sphereface loss is right but arcface loss ? error

[seanM29]

i meet same problem on mnist(resnet 50, arcface loss), have you solved this problem? @damvantai