how to update v_i about pfedhn HOT 6 CLOSED

Hey,
Please note that v_i is part of hnet.parameters(), so the update is consistent (mathematically) with the one specified in Alg. 1 of the paper.

from pfedhn.

Thanks for your prompt response.

Yes, I understood that v_i is part of hnet.parameters().

What does hnet_grads (line 172 in the code) imply for?

• Whether it is \delta_phi \theta^T \delta \theta for updating \phi?
• And it is \delta_vi \phi^T \delta_phi \theta^T \delta \theta for updating v_i?

I am still confused about where the term \delta_vi \phi^T goes in the code if the aforementioned is not the case. Since you have all in the hnet_grads, and use it to update both \phi and v_i.

Thanks,

from pfedhn.

And one more question, in models.py, lines 54 and 55,
emd = self.embeddings(idx)
features = self.mlp(emd)
"features" and "emb", which one is v_i ?

from pfedhn.

hnet_grads contains the gradients for all the parameters of the HN (v and phi). The embeddings v are just another layer in the HN, so there's really no difference between v and phi.

For your second question, v is the embedding weights self.embeddings.weight. emb from emd = self.embeddings(idx) is the equivalent of v_i from the paper.

from pfedhn.

It is clear to me. Thanks a lot for your answers! I closed the issue.

from pfedhn.

@AvivNavon You mentioned that "hnet_grads contains the gradients for all the parameters of the HN (v and phi). The embeddings v are just another layer in the HN, so there's really no difference between v and phi."

So, it means that \delta_vi \phi^T = 1 ?

from pfedhn.