Margin in NCA loss

[EvelynChee]

incremental_learning.pytorch/inclearn/lib/losses/base.py

Lines 72 to 74 in 0d25c2e

	margins = torch.zeros_like(similarities)
	margins[torch.arange(margins.shape[0]), targets] = margin
	similarities = scale * (similarities - margin)

Based on the equation in the paper, shouldn't line 74 be similarities - margins instead? Currently it seems to remove the margin for all classes instead of the only the target class?

[arthurdouillard]

Line 73 does that, margins is zero for all negative class, and of value margin for the target class.

[EvelynChee]

@arthurdouillard yes, margins has that but in line 74, the value margin is deducted instead of margins?

[arthurdouillard]

Oh yes, you're definitely right!

Thanks for catching that, I'll fix it.