Question about EWC Implementation

[yunboouyang]

Hello developers! I looked at the implementation of EWC and got confused about computing the diagonal of the empirical Fisher Information Matrix. In

GradientEpisodicMemory/model/ewc.py

Line 88 in 34c6b8e

pg = p.grad.data.clone().pow(2)

the square of the sum of gradients of all data points in the memory are computed, but this is not equal to the sum of squares of gradients. The latter should be used to compute Empirical Fisher, see https://wiseodd.github.io/techblog/2018/03/11/fisher-information/. Could you confirm this?