add Google Summer of Code [GSOC] doc page (#117)

* add GSOC proposal

* relax atol, rtol GATv2

Author: CarloLucibello

docs/make.jl

@@ -24,6 +24,7 @@ makedocs(;
                 "Utils" => "api/utils.md",
                ],
               "Developer Notes" => "dev.md",
+              "Summer Of Code" => "gsoc.md",
             ],
 )
 

docs/src/gsoc.md

@@ -0,0 +1,36 @@
+# Graph Neural Networks - Summer of Code
+
+Graph neural networks (GNN) are deep learning models well adapated to data  rapresented by graphs along with feature vectors associated to nodes and edges. 
+GNNs are a growing are of research and find many applications in complex networks analysis, relational reasoning, combinatorial optimization, molecule generation, and many other fields. 
+[GraphNeuralNetworks.jl](https://github.com/CarloLucibello/GraphNeuralNetworks.jl) is a pure Julia package for GNNs already equipped with plenty of features such as many common graph convolutional layers, CUDA support and graph batching for fast parallel operations. There are a number of ways by which the package could be improved:
+
+- **Adding graph convolutional layers (duration: 175h, expected difficulty: easy to medium):**
+While we implement a good variety of graph convolutional layers, there is still a vast zoology to be implemented yet.
+
+- **Adding models and examples (duration: 175h, expected difficulty: medium):**
+As part of the documentation and for bootstrapping new projects, 
+we want to add fully worked out examples and applications of graph neural networks.
+
+- **Adding graph datasets (duration: 175h, expected difficulty: easy):**
+. Provide julia friendly wrappers for common graph datasets in [`MLDatasets.jl`](https://github.com/JuliaML/MLDatasets.jl).
+
+- **Supporting heterogeneous graphs (duration: 175h, expected difficulty: hard):**
+In some complex networks, the relations expressed by edges can be of different types.
+We need to implement an heteroeneous graph type and implement convolutional layers supporting them. 
+
+- **Training on very large graphs (duration: 175h, expected difficulty: medium to hard):**
+Graph containing several milions of nodes are too large for gpu memory. Mini-batch training si performed on subgraphs, as in the GraphSAGE algorithm.
+
+- **Supporting temporal graph neural networks (duration: 175h, expected difficulty: hard):**
+We aim at implementing temporal graph convolutions for time-varying graph and/or node features.
+
+- **Improving perfomance using sparse linear algebra (duration: 175h, expected difficulty: medium to hard):** 
+Many graph convolutional layers can be expressed as non-materializing algebraic operations involving the adjacency matrix instead of the slower and more memory consuming gather/scatter mechanism. We aim at extending as far as possible and in a gpu-friendly way these *fused* implementation.
+
+**Recommended skills:** Familiarity with graph neural networks and Flux.jl.
+
+**Expected results:** New features added to the package along with tests and relevant documentation.
+
+**Mentors:** [Carlo Lucibello](https://github.com/CarloLucibello) (author of [GraphNeuralNetworks.jl](https://github.com/CarloLucibello/GraphNeuralNetworks.jl).For linear algebra, co-mentoring by [Will Kimmerer](https://github.com/Wimmerer) (lead developer of [SuiteSparseGraphBLAS.jl](https://github.com/JuliaSparse/SuiteSparseGraphBLAS.jl)).
+
+**Contact:** Feel free to contact us on the [Julia Slack Workspace](https://julialang.slack.com/) or by opening an issue on [GraphNeuralNetworks.jl](https://github.com/CarloLucibello/GraphNeuralNetworks.jl).

test/layers/conv.jl

@@ -98,7 +98,7 @@
         for heads in (1, 2), concat in (true, false)
             l = GATConv(in_channel => out_channel; heads, concat)
             for g in test_graphs
-                test_layer(l, g, rtol=1e-4,
+                test_layer(l, g, rtol=1e-3, atol=1e-3,
                     outsize=(concat ? heads*out_channel : out_channel, g.num_nodes))
             end
         end
@@ -114,7 +114,7 @@
         for heads in (1, 2), concat in (true, false)
             l = GATv2Conv(in_channel => out_channel; heads, concat)
             for g in test_graphs
-                test_layer(l, g, rtol=1e-4,
+                test_layer(l, g, rtol=1e-3, atol=1e-3,
                     outsize=(concat ? heads*out_channel : out_channel, g.num_nodes))
             end
         end