I am trying to build a face recognition application and I have seen implementations such as dlib. I would like to build a siamese net, and my doubts are about the architecture.
Since I am supposed to compare positive, negative and original samples, is the architecture supposed to be having 2 networks for both, if so how can I later convert it to a single network for inference, ie. getting embeddings for a single sample and comparing it with stored embeddings instead of images.
As I think 2 networks is not the way to go, is it supposed to be a special kind of loss function which accumulates embedding data for 3 consecutive samples(for triplet loss) and then calculate the loss and backpropogate(after averaging for some such pairs in a minibatch)
If none of this, how can I achieve such thing.