GeometricFlux.jl：Flux 上的幾何深度學習¶
GeometricFlux.jl: Geometric Deep Learning on Flux
Speaker: Yueh-Hua Tu（杜岳華）¶
Why graph neural network?¶
Nowadays AI¶
- Computer vision
- Natural language processing
- Speech
- Game, including cheese game, real-time strategy (RTS)
Many scientific data lies in non-Euclidean space¶
- Social networks in socail science
- Traffic networks
- Gene regulatory networks in biology
- Molecular structure in chemistry
- Knowledge graph
- 3D object surfaces in computer graphics
Graph¶
They can be represented as a graph structure.
Human gene coexpression network¶
Computer graphics¶
Knowledge graph¶
Graphs are more representative than vectors¶
Graphs are discrete approximation to manifold¶
Geometric Deep Learning¶
Graph + Deep Learning = Powerful¶
GeometricFlux¶
Aims to¶
- extend Flux deep learning framework in Julia
- support of CUDA GPU with CuArrays (other interfaces are welcome)
- integrate with existing JuliaGraphs ecosystem
Graphs are usually sparse¶
I worked on SparseArrays and CuArrays.CUSPARSE
Layers¶
- Convolution layers
- MessagePassing
- GCNConv
- GraphConv
- ChebConv
- GatedGraphConv
- GATConv
- EdgeConv
- Meta (WIP)
- Pooling layers
- GlobalPool (WIP)
- TopKPool (WIP)
- MaxPool (WIP)
- MeanPool (WIP)
- sum/sub/prod/div/max/min/mean pool
- Embedding layers
- InnerProductDecoder
Models¶
- VGAE
- GAE
Compatible with layers in Flux¶
In [ ]:
## Model
model = Chain(GCNConv(g, num_features=>1000, relu),
GCNConv(g, 1000=>500, relu),
Dense(500, 7),
softmax)
Use it as you use Flux¶
In [ ]:
## Loss
loss(x, y) = crossentropy(model(x), y)
accuracy(x, y) = mean(onecold(model(x)) .== onecold(y))
## Training
ps = Flux.params(model)
train_data = [(train_X, train_y)]
opt = ADAM(0.01)
evalcb() = @show(accuracy(train_X, train_y))
Flux.train!(loss, ps, train_data, opt, cb=throttle(evalcb, 10))
Construct layers from SimpleGraph/SimpleWeightedGraph¶
In [ ]:
g = SimpleGraph(5)
add_edge!(1, 2); add_edge!(3, 4)
GCNConv(g, num_features=>1000, relu)
Use Zygote (with CPU)¶
GeometricFlux with Zygote on GPU is not available. There are some issues to work on...
For GPU, I will getting Tracker work first.
Use message passing scheme¶
$$
m_v^{t+1} = \sum_{u \in N(v)} M_t(h_v^t, h_u^t, e_{uv}) \\
h_v^{t+1} = U_t(h_v^t, m_v^{t+1})
$$
$h_v^t$: vertex features, $e_{uv}$: edge features, $N(v)$: neighbors of vertex $v$
Message passing scheme¶
Performance¶
- Multithreaded scatter function
- Scatter function on GPU (WIP)
Some work to do¶
- Support CUDA/CUSPARSE
- Accept edge list dynamically
- More layers/models
- Datasets inetegration
- Threading on CPU (in v1.3!)
- Sparse array computation optimization