Hey Santiago,

Yes, it is possible to implement the triplets generation part inside the network for each batch. It is called Online triplet Mining. It is indeed more efficient in terms of memory and speed. Check out this link: https://omoindrot.github.io/triplet-loss