Monitor training process with batch hard

[FSet89]

When using batch hard, is it correct to say that the hardest negative distance should increase and the hardest positive distance should decrease during training? Is it a good method to monitor the training process?

[paweller]

Yes, it is true that the inter-class (or negative) distances should increase while the intra-class (or positive) distance should decrease during training. The triplet loss aims to do exactly that.

Regarding the monitoring process you can use the mean inter-class or intra-class distance. However, as those can differ a lot from project to project, you might want to rather use other metrics. These could be a triplet error rate (number_of(d(A,P)-d(A,N) > 0) / number_total_triplets) or the AUC-ROC metric or simply the L2 distance.