Closed wenjie710 closed 2 years ago
Do you perform L2 normalize after GeM pooling during training? What is the range of the distance in GCL loss, [0, +inf] or [0, 2]?
Dear @wenjie710 At first sight, your code looks fine to me, but I haven't debugged it. I do not perform the nan check though, since I did not get nan values in my experiments. I do perform L2 normalization after GeM pooling, and therefore the euclidean distance has to be necessarily between 0 and 2, the maximum possible Euclidean distance between vectors of norm = 1. Please let me know if the L2 normalization solves your issue.
Thanks a lot. Normalization solves the issue.
Great to hear! I'm closing this, but feel free to re-open or drop me an email if you have further questions.
Hi, I am trying to reproduce the results in the paper using ResNet152 as backbone and GCL loss. However, I get nan error during training. Here is my implementation for GCL loss.
Could you please give me some advice? Do I get something wrong?