Understanding MetaSGD

[emit33]

I'm struggling to understand the use of MetaSGD and hoped you might help me. To my understanding, in short, MetaSGD proposes learning parameter initialisations, where as MAML proposes learning parameter intialisations as well as per-parameter learning rates (where here the parameters are the entries in a modulation vector). The paper mentions use of MetaSGD, while helpers.inner_loop says that it performs MAML, but it seems to me that the modulations are only initialised once at the very beginning. Could you tell me if I've misunderstood the code and/or the references?

Thanks a lot!