Dimension mismatch in GCN with DisjointLoader and node_level=True

[prevay]

I have a custom dataset with multiple graphs that I want to run a node classification task on using a GCN. As a proof of concept I start with the Citation GCN example, and instead read in one of the included datasets with multiple graphs (like QM7) and load the data using DisjointLoader with node_level=True. However, this results in a dimension mismatch when applying the second convolutional layer. Specifically, I get the following error:

Dimensions must be equal, but are 1 and 289 for '{{node gcn_1/gcn_conv_1_2/MatMul_1}} = MatMul[T=DT_FLOAT, grad_a=false, grad_b=false, transpose_a=false, transpose_b=false](inputs_3, gcn_1/gcn_conv_1_2/MatMul)' with input shapes: [289,1], [289,14].

This particular example is for a single graph (batch_size=1) with a (16,16) SparseTensor for the adjacency matrix, a (256,1) ndarray for edge features and a (16,) ndarray for node features.