About the denominator of classification loss about actionformer_release HOT 4 CLOSED

Good catch! The denominator of the classification loss (Eq. 7 of our tech report) should be the number of positive samples. We will update the tech report soon.

In practice, self.loss_normalizer saves a moving average of the number of positive samples. The classification loss is summed over all samples, where the regression loss is summed over positive samples. Both are divided by self.loss_normalizer and are then added up.

from actionformer_release.

Thanks for the reply! The moving average normalizer is a really good trick for stabilizing small batch training!

I still want to have a further discussion about the classification loss. I guess it should be divided by the number of all samples, just like in the original focal loss, instead of only positive samples. However, I also find that if in this way (divided by a moving average of the number of all samples), the classification loss value would be small and the final detection result is poor.

So I wonder do you have any specific motivation for choosing such implementation, or just because it is a good weight for balancing the regression loss and classification loss. Please correct me if I am wrong. Thanks!

from actionformer_release.

Nice to have this discussion. The modern version of the focal loss (RetinaNet) uses a similar implementation, where both the classification loss (focal) and the regression loss (smoothed L1) are divided by the moving average of number of positive samples. See the code here in Detectron2. I think the choice here is mostly empirical.

from actionformer_release.

Good to know that! Now it is clear.

Have a nice day!

from actionformer_release.