Has the EWC of the paper been fully implemented?

[wangxiaodong1021]

I am glad to see the implementation of the work of EWC, but I feel trouble with you to achieve the EWC.
Why is the call to tf.nn.l2_loss when calculating EWC?

I was wondering if this was temporary replacement of EWC with L2 regularization because I did not see where to calculate the Fisher information matrix.

[wangxiaodong1021]

All right, I just know that the Fisher is square(grad). But, in the paper, the EWC function is defined as λ/2*sum(F(θ-θA*)^2). In this code, EWC is implemented by λ/2*sum((F*(θ-θA*))^2). I'm not sure if I understand it is reasonable.