I'm trying to train a Siamese network for face Verification and eventually I came across the Contrastive Loss method for embedding vector distancing (kinda... I guess...).
At the end, the model needs to yield '1' if two images has the same person in it, else '0'.
I though maybe I should first train the CNN to yield correct vectors with the "Contrastive Loss", and only after training was finished, add a FC layer with 1 output, and train again with the "Binary Cross Entropy", but (and thats the question) Is that mandatory? or I could train the model with both losses simultainously?
Does the steps on training the model has effects on the training?
