Added GMMConv

[itsmohitanand]

Added GMMConv from the paper: Geometric deep learning on graphs and manifolds using mixture model CNNs

[CarloLucibello]

Tests are missing

[codecov]

Codecov Report

Merging #147 (65488a1) into master (04d026b) will decrease coverage by 0.08%.
The diff coverage is 70.58%.

@@            Coverage Diff             @@
##           master     #147      +/-   ##
==========================================
- Coverage   85.61%   85.52%   -0.09%     
==========================================
  Files          15       15              
  Lines        1251     1285      +34     
==========================================
+ Hits         1071     1099      +28     
- Misses        180      186       +6     

Impacted Files	Coverage Δ
src/layers/conv.jl	78.70% <70.58%> (-0.82%)	⬇️
src/msgpass.jl	73.80% <0.00%> (+9.52%)	⬆️

---

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 04d026b...65488a1. Read the comment docs.

[itsmohitanand]

Added the changes.

• Added the learnable parameter in the doc
• Changed the name from u to e in view of the package convention
• As a convention, last dimension as num_edges.

Still need to add the test.

[itsmohitanand]

Test added as well

[CarloLucibello]

A first test failure could be solved by adding

(l::GMMConv)(g::GNNGraph) = GNNGraph(g, ndata=l(g, node_features(g), edge_features(g)))

[itsmohitanand]

I understand the error 1, this is that if my input is Float32 still my output is Float64, will see why this is happening. Could not understand the second test.

[CarloLucibello]

Suggested change

	b = bias ? Flux.create_bias(ones(out), true) : false
	b = bias ? Flux.create_bias(mu, true, out) : false

This should fix the Float64 issue

[itsmohitanand]

The problem is here. Both w and mu are Float32 before this, but it changes to Float64 at this point. How exactly @. functions? w = @. -0.5 * (w - mu)^2

[itsmohitanand]

I don't know if this is the most elegant solution, but I had to change the code at two different points.

• w = @. -0.5 * (w - mu)^2 goes to w = @. - ((w - mu) ^ 2)/ 2
• m = 1 / d .* m goes to m = m./reshape(d, (1, g.num_nodes))
  The problem is if in degrees say d=[1,2,1] is operated to 1/d then it converts to Float 64, depending on the system. Same with 2 and 0.5 I guess.

[CarloLucibello]

Could not understand the second test.

Me neither. Seems to be a Zygote error due to a not supported try/catch. But we already have @assert and @warn in other layers and they don't seem to cause problems. Maybe try to simplify the forward a bit and test locally with something like

l = GMMConv((2,2)=>2)
g = rand_graph(5, 10, ndata=rand(Float32,2,5), edata=rand(Float32,2,10))
gradient(() -> sum(l(g, g.ndata.x, g.edata.e)), Flux.params(l))  

[itsmohitanand]

Could not understand the second test.

Me neither. Seems to be a Zygote error due to a not supported try/catch. But we already have @assert and @warn in other layers and they don't seem to cause problems. Maybe try to simplify the forward a bit and test locally with something like

l = GMMConv((2,2)=>2)
g = rand_graph(5, 10, ndata=rand(Float32,2,5), edata=rand(Float32,2,10))
gradient(() -> sum(l(g, g.ndata.x, g.edata.e)), Flux.params(l))  

It is probably related to the string interpolation error in gradient. But changing this
"Pseudo-cordinate dim $(size(e)) does not match (ein=$(ein),num_edge=$(g.num_edges))"
to
"Pseudo-cordinate dim does not match (ein, num_edge))"
runs everything smoothly

[itsmohitanand]

NANs possibly because dividing by indegree, instead changed the aggregator method to mean. This should potentially get. rid of the NAN values

[CarloLucibello]

Great, impressive works, thanks!

[itsmohitanand]

Great, impressive works, thanks!

Thanks a lot for all the help! Will try some other conv layer this weekend. :) And then will see about temporal GNN.