Update loss.py

[banjuanshua]

Fix clc loss nan bug. Sum of cls_loss should divide total number of pos_neg

[SCoulY]

Good job

[pengfeidip]

@banjuanshua Have you successfully run the code?

[ridhajuneja123]

This is also giving nan loss how the above solution solve the problem can u explain

[noahcao]

For the best, you should also add clamp in the focal_loss in case the sigmoid produce inf value:

def focal_loss_alt(self, x, y):
        '''Focal loss alternative.

        Args:
          x: (tensor) sized [N,D].
          y: (tensor) sized [N,].

        Return:
          (tensor) focal loss.
        '''
        alpha = 0.25

        t = one_hot_embedding(y.data.cpu(), 1+self.num_classes)
        t = t[:,1:]
        t = Variable(t).cuda()

        xt = x*(2*t-1)  # xt = x if t > 0 else -x
        pt = (2*xt+1).sigmoid()

        w = alpha*t + (1-alpha)*(1-t)
        logit = pt.clamp(self.eps, 1.-self.eps)
        loss = -w*logit.log() / 2
       
        return loss.sum()

This should be better, where we could set self.eps=1e-7

[zjysteven]

Have you trained a satisfying model with your modification? In my case, dividing cls_loss with pos_neg will give a very small cls_loss (e.g. around 0.0020) since the number of background anchors is so large. However, after training, the classification outputs were all near 0 and I think it is because such a small cls_loss prevented the model from learning.

Also, other implementations below both uses num_pos as the normalizer. So I don't think your PR is actually helpful.
https://github.com/fizyr/keras-retinanet/blob/a8bbd6535655602a84f194b1557be1ed5180fa71/keras_retinanet/losses.py#L61-L65
https://github.com/yhenon/pytorch-retinanet/blob/bb11485e410dba9da3578de5b2cb36d9cd6fdd03/retinanet/losses.py#L92

[twmht]

@zjysteven

what is the training cls loss when you use only num_pos as the normalizer?

[zjysteven]

@twmht Sorry it's been a long time so I can't remember what exactly the value of that is. But I have confirmed with the authors of several other implementations, and it is correct to divide the cls loss by only num_pos. It is also the case in Facebook's Detectron2, so you can be assured to do so. However, according to my own experience, this repo may have some other subtle implementation errors so it is not guaranteed that you get desirable results with this repo.