xiangli13/circle-loss

tf2.0 implementation of circle loss

circle-loss's Issues

is it some fault?

circle-loss/circle_loss.py

Line 33 in 13aa7ab

margin_n = -self.margin

your code : margin_n = -self.margin, but in the paper,it is self.margin?
your code : logit_p = tf.reshape(self.scale * alpha_p * (sim_p - margin_p), (-1, 1)), but in the paper, it is
-self.scale

consult not the real issue

hi committer, thanks for sharing your work！
but confiused with the comment of the CircleLoss
"Inputs is a list of three tensors, the first one contains the positive feature samples, the second contains the negative feature samples, and the third contains the query feartures. "

The features I got from the model and surely with the labels(just like , 1,2,3... not the one_hot method). these are only two variabels here , inputs are array include three variables.

how could i use this inputs in my classification task.

How to handle the different number of positive and negative samples?

Thanks for sharing the code!

I noticed that the implementation supports the calculation between 1 positive pair and 1 negative pair.
If I have 1 positive pair and N negative pair, I am wondering how to handle the loss more effectively?

Thanks!