Open qiyang77 opened 5 years ago
you'd better refer the Large-Margin Softmax Loss for Convolutional Neural network, which explains the lambda's meanings.
but, here,the code. l, f, ff is constant.which isn't right. l = 0. f = 1.0/(1.0+l) ff = 1.0 - f
the l should depends at the global_steps.
Hello, I change the "l" crosspond to the training step, and let angular-softmax with a very small part at beginning and about 0.1 ratio at the end. however, it seems strange, the loss decrease normally at beginning and start increase at some training step, with a-softmax increase. and the acc decrease too, have anyone meet this problem? If anyone have some trick about the raito decay? ps: I have fixed the code to avoid gradient explosion problem.
hello, I am a bit of confuse about reading this code about calculating A-softloss