I have read the paper about circle loss for neural networks. I may have missed something, but I didn't find the way to compute the positive similarity and the negative similarity in the case of circle loss.
The paper gives the equations for AM-SoftMax and Triplet loss, that can be computed from a unified loss function:
For example, in the case of AM-SoftMax, the positive and negative similarities are computed by the following way,
w being the weights of the last fully connected layer, and
x being the inputs of this last layer:
So my questions are:
Should I compute the similarities for circle loss by the same way the paper compute them for AM-SoftMax ?
Should I compute the forward pass of the last fully connected layer with a standard product:
output[j] += weight[i][j] * input[i], followed by a standard SoftMax function ? Or by another way ?