Excuse me, maybe an stupid question. I am new in DL, because I have a few training data for a classification project I proposed to use Siamese network (It allows few-shot learning). After watching deeplearning.ai videos I got confused:
Are these correct?
Siamese is trained in two steps:
- First, training step, using the triplet loss to learn the model parameters (https://www.youtube.com/watch?v=g3UGGAEbw_Q)
- Next, recognition step, add a Sigmoid activation to get a binary classifier (https://www.youtube.com/watch?v=0NSLgoEtdnw). In this case, I must retrain the net, in this case using another loss, ex. Hinge Loss?