Closed devmotion closed 3 years ago
The bracket thing is my own derivation to speed up the root finding algorithm (I haven't seen a discussion of it anywhere else), the bugfix (removal of norm(w, 2)
) can be checked by looking at eq. (22) in the original paper https://arxiv.org/pdf/1505.05770.pdf and plugging in the parameterization of z|| with alpha below the equation (we solve eq. (22) for z = z| + z||).
There are multiple references to the original paper in the source code but maybe not to eq. (22) explicitly at this line.
I had a look at the planer layer and found a bug in the implementation of the inverse.
However, one can also rewrite the forward pass in a simpler and, it seems, numerically more stable way. To keep the PR small, I don't include these possible changes here and just try to fix the test errors.