You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi!
I have a question for the loss re-weighting.
According to my understanding, image with higher k should be assigned a lower loss weight since it is easier to predict with more information disclosed. As in Fig. 3, cat image has area k, and dog image 1-k. The computed loss weight is w(1-k) for cat and w(k) for dog. The higher k, the higher w(k).
But in the code, it seems the computed weight w(k) is assiged to image with ratio k, other than the opposite.
I am wondering which one is correct? How to understand the difference here?
The text was updated successfully, but these errors were encountered:
Hello, thank you for your question.
It's the other way around: the main image (k>1/2) should be assigned more weight in classification than the other "minor" image (1-k<1/2). Generally, in mixup/cutmix papers, this weight is exactly the ratio k. We found that in MixMo, taking a "power-law" formulation works better to improve diversity. See more information in Appendix 6.1. And sorry if Fig 3 was misleading. So to sum up, the implementation is correct and consistent with what was described in the paper. Hope this helps.
Hi!
I have a question for the loss re-weighting.
According to my understanding, image with higher k should be assigned a lower loss weight since it is easier to predict with more information disclosed. As in Fig. 3, cat image has area k, and dog image 1-k. The computed loss weight is w(1-k) for cat and w(k) for dog. The higher k, the higher w(k).
But in the code, it seems the computed weight w(k) is assiged to image with ratio k, other than the opposite.
I am wondering which one is correct? How to understand the difference here?
The text was updated successfully, but these errors were encountered: