R1 loss - called with all params

[moshebeutel]

Hi , reading your paper I understood that the first custom loss is regularized to minimize personalized parameters change. It seems that in line 112 in ours.py the diffs are calculated using all model.parameters() and w_glob and not just the personalized. at the end of iteration only personalized parameters are updated but I do not think this is equivalent to computing the diff on the personalized parameters as described in the paper. Am I missing something?
Thanks 🙏 Moshe

[moshebeutel]

Same holds for the second loss / shared parameters

[xiyuanyang45]

Then if there's any alternatives for only calculating updates for partial selected params?
I think that code is an available approach for implementation, welcome to discuss any other possible implementations of this algorithm