Pytorch-Additive_Margin_Softmax_for_Face_Verification
Pytorch Implementation of AMSoftmax with label smooting
Additive Margin Softmax for Face Verification
Feng Wang, Weiyang Liu, Haijun Liu, Jian Cheng
https://arxiv.org/pdf/1801.05599.pdf
from AM_Softmax import AM_Softmax_v1 , AM_Softmax_v2
features = Model(input)
criterion_xent = AM_Softmax_v1()
Loss = criterion_xent(features, labels)
or
criterion_xent = AM_Softmax_v2()
Loss = criterion_xent(features , labels, model.module.classifier )
[Only for V_1] Since weights are in Loss function only, add these to the model optimizer as well.
for key, value in criterion_xent.named_parameters():
params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]
optimizer = torch.optim.Adam(params)
If you don't want label smoothing : remove the line
targets = (1 - self.epsilon) * targets + self.epsilon / self.num_classes