Labels: background as 1, foreground as 0?

[yt2639]

Hi guys, thanks for your work and sharing the code. I have a question about the labels input to calculate the loss. So I understand it as if we have multi-class detection problem, say 5 categories, then the foreground would be 0,1,2,3,4 and the background will be 5. So similarly, if we only have 1 class, then foreground would be 0, background would be 1.

I was just wondering whether this "fg-0 bg-1" has been flipped (as in "fg-1 bg-0") in calculating the loss? Cuz I saw from the vadacore.ops.sigmoid_focal_loss, specifically in sigmoid_focal_loss_cuda.cu file, it wrote

__global__ void SigmoidFocalLossForward(const int nthreads,
                                        const scalar_t *logits,
                                        const int64_t *targets,
                                        const int num_classes,
                                        const float gamma, const float alpha,
                                        const int num, scalar_t *losses) {
  CUDA_1D_KERNEL_LOOP(i, nthreads) {
    int n = i / num_classes;
    int d = i % num_classes;  // current class[0~79];
    int t = targets[n];       // target class [0~79];

    // Decide it is positive or negative case.
    scalar_t c1 = (t == d);
    scalar_t c2 = (t >= 0 & t != d);

And I guess this int d = i % num_classes; // current class[0~79] is where the labels are flipped (so labels become bg-0 fg-1)?

The reason why I have this question is when I look at the loss, if the labels aren't flipped, it doesn't make sense. For the simplest case, Binary Cross Entropy loss, it should be

loss = - [y log(p) + (1-y) log(1-p)]

Minimizing the loss is equivalent to maximizing y log(p) + (1-y) log(1-p). So here, when y=1, we maximize p; when y=0, we maximize 1-p i.e. minimize p. And so here, if the input labels are in "bg-1 fg-0", we should make it "bg-0 fg-1". Is this correct?

Thanks!

[yt2639]

@hxcai @mileistone Could you answer this question? Thanks!