Skip to content

Simplify graph code in conv layer

Vicentini Filippo requested to merge github/fork/femtobit/feature/feedforward into feature/feedforward

Created by: femtobit

As far as I understand, the following part of the code in the Convolutional class is supposed to store, for each node i, all nodes with distance of at most dist_ to i in neighbours_[i]: https://github.com/netket/netket/blob/67fe7e62c93abf4280d85c16023563868b3e41bd/NetKet/Machine/conv_layer.hpp#L115-L137

I found the logic there a bit hard to read an rewrote it in terms of a breadth-first search started from each node, which I find easier to understand. (It could also be a bit faster, but I did not check this.) The BFS code can potentially also be useful elsewhere.

If this looks good to you, accepting this PR will merge the changes into the branch for the feedforward PR #46 (so this is a PR on a PR, essentially).

Merge request reports