Question about regularization loss

[willard-yuan]

Hi @lucasb-eyer, I have done a experiment on an instance image retrieval. The loss is converged well, and the batch-p@K goes to almost 100%. I used the model to tested it on Oxford Building dataset, and the mAP is 0.6~0.7. It's not a promising result, I'm curious about whether the network is overfit? I found no regularization loss in the codes.