When I read the prototypical paper Prototypical Networks for Few-shot Learning, I understand in Eq. 1 that I should train $f_\phi$, which takes as input $x_i$, which is already an embedding of an image. So, I understand from the paper that I should not train my backbone further, but only a new network that takes as input a first set of embeddings.
But in every github repo I find that implements prototypical networks, they train the whole network (using the loss defined just after Eq.2).
Did I misunderstood something?