Use of the margin

[gombru]

Hello,
I'm successfully using your tripletlosslayer.py to train a triplet net, but I have some doubts:

• The self.margin parameter is only used in the forward() function to compute the loss, but not in backward. As I understand, in Caffe the gradient should be set in bottom[].diff in the backward() function to do the backpropagation. So as the code is, the margin does not affect the training, but only the loss display, right?
• I see than in setup you set self.a to 1, which you later use in backward for gradient computation. Why do you use that parameter and set it to 1?

[luhaofang]

Hi, I think the margin is the hyper parameter to modify the positive triplet case, you may find that some triplet samples DON'T need to do backpropagation.

Please ignore the parameter self.a.

[gombru]

OK. So as I see in the code, the margin value is only used to omit from backpropagration those triplets fulfilling the margin, but has no influence in those triplets that are backpropagated, right?