Use GCN to classify Mnist
The implementation refers to "Semi-supervised classification with graph convolutional networks" and "Hierarchical graph representation learning with differentiable pooling".
Each digit, a 28x28 matrix, is viewed as a 28-node graph. Each node is a 28d vector. We do this just to study GCN.