SGConv

[rbSparky]

Adding SGConv from the paper Semi-Supervised Classification with Graph Convolutional Networks.

[codecov]

Codecov Report

Merging #154 (97bb11c) into master (104b6b9) will decrease coverage by 0.07%.
The diff coverage is 65.00%.

❗ Current head 97bb11c differs from pull request most recent head 28ee1de. Consider uploading reports for the commit 28ee1de to get more accurate results

@@            Coverage Diff             @@
##           master     #154      +/-   ##
==========================================
- Coverage   85.52%   85.45%   -0.08%     
==========================================
  Files          15       14       -1     
  Lines        1285     1299      +14     
==========================================
+ Hits         1099     1110      +11     
- Misses        186      189       +3     

Impacted Files	Coverage Δ
src/layers/conv.jl	77.23% <65.00%> (-1.47%)	⬇️
src/GNNGraphs/gnngraph.jl	76.00% <0.00%> (-4.00%)	⬇️
src/GNNGraphs/sampling.jl	100.00% <0.00%> (ø)
src/deprecations.jl
src/GNNGraphs/transform.jl	96.74% <0.00%> (+0.03%)	⬆️
src/GNNGraphs/query.jl	93.02% <0.00%> (+0.04%)	⬆️
src/GNNGraphs/convert.jl	90.40% <0.00%> (+0.31%)	⬆️
src/GNNGraphs/utils.jl	84.72% <0.00%> (+10.23%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2801b51...28ee1de. Read the comment docs.

[rbSparky]

What I understand from the paper is that we create a layer, which has the same feature propagation with the normalised adjency matrix step as GCNConv, but for the k hops, we propagate k times (which can be done with a loop?), and then in the end, we return the return value without applying the activation function.

Can you confirm if I have the right idea? I will try to test it and then make the next commit.

[CarloLucibello]

You put the wrong reference in the opening post, the correct one is https://arxiv.org/pdf/1902.07153.pdf

[CarloLucibello]

What I understand from the paper is that we create a layer, which has the same feature propagation with the normalised adjency matrix step as GCNConv, but for the k hops, we propagate k times (which can be done with a loop?), and then in the end, we return the return value without applying the activation function.

Yes that's correct. Instead of copying over the implementation of GCNConv forward pass, we could actually store a GCNConv with no activation inside of the SGConv and use that K times in the forward pass.

[rbSparky]

Yes that's correct. Instead of copying over the implementation of GCNConv forward pass, we could actually store a GCNConv with no activation inside of the SGConv and use that K times in the forward pass.

Makes sense, I'll do that.

EDIT: Previous comment(which I removed since I wasn't quite sure about it):
For SGConv, after removing the non linearity, we have the equation but we can then further collapse the linear transformation into one of such , wouldn't using GCNConv k times implement the first equation?

You put the wrong reference in the opening post, the correct one is https://arxiv.org/pdf/1902.07153.pdf

Oh yes, my apologies, I've fixed that now.

[CarloLucibello]

ouch sorry, I was mistaken, applying the GCNConv multiple times is not equivalent to SGConv since we don't want to multiple by the weights multiple times but only once. Sorry for derailing you!

[rbSparky]

I think I should now move on to testing this?, if there are any changes to be made please let me know. Should I write some tests similar to the ones for other layers, since I'm not very sure how the specifics of this would be tested though(like the final result for example, or if we even need to test that). What do you suggest?

[CarloLucibello]

The implementation looks good. For testing you can follow the example of the other layers. We test output size, equality between cpu an gpu outputs, differentiability.

[rbSparky]

I've added the tests, is there anything else that needs to be done before merging?

[CarloLucibello]

thanks for this!