Nonconvergence caused by Generalized-mean (GeM) Pooling？

[Hellomodo]

Here are some tips for debugging the nonconvergence problem:

1. Replace the self.p in GeneralizedMeanPooling as a constant 3. Train the model again, and see if the nonconvergence happens?
2. Replace the self.p in GeneralizedMeanPooling as a constant 1. Train the model again, and see if the nonconvergence happens? In this way, the GeneralizedMeanPooling degrades to a normal global average pooling operation.

[Margrate]

Thank you for your reply. When I use normal global average pooling, it can converge normally. Did you compare the effects of Weighted Regularization Triplet (WRT) loss and Batch Hard triplet loss?

[Hellomodo]

If it can converge normally when you use normal global average pooling, then it should also converge normally when you replace the self.p in GeneralizedMeanPooling as a constant 3. Because in this way, the GeneralizedMeanPooling is just a tradeoff between global average pooling and global max pooling. Or you can try whether your model can converge normally with global max pooling?

Weighted Regularization Triplet (WRT) performs a little better than Batch Hard triplet loss in our experiments.

[Margrate]

In my experiments, softmax loss + center loss + Batch Hard triplet loss is better than softmax loss+ center loss + Weighted Regularization Triplet (WRT) loss. Maybe my backbone is different from you.

[tomFoxxxx]

In my experiments, softmax loss + center loss + Batch Hard triplet loss is better than softmax loss+ center loss + Weighted Regularization Triplet (WRT) loss. Maybe my backbone is different from you.

That's normal.