NeighborSampler for k-hop sampling

❓ Questions & Help

I have a hard time understanding NeighborSampler even with small examples. It doesn’t make sense in most cases.

l = np.array([[0, 1], [1, 4], [1, 5], [1, 6], [0, 2], [2, 7], [0, 3], [3, 8], [3, 9], [3, 10], [3, 11]]).transpose()
r = np.array([l[1, :], l[0, :]])
edge_index = torch.from_numpy(r)

train_loader = NeighborSampler(edge_index,
                               sizes=[2, 2], batch_size=1)

for batch_size, n_id, adjs in train_loader:
    print("###")
    for edge_ind, e_id, size in adjs:
        print(e_id)

gives me:

tensor([ 4,  0, 10,  8,  3,  1])
tensor([6, 0])

It looks like it jumps over edges?

How can I use it for k-hop sampling?

[rusty1s]

The NeighborSampler example looks correct to me. It samples two edges in the last layer for node 0: to nodes 1 and 3. in the first layer, it samples 6 edges for nodes 0, 1 and 3: to nodes 1, 4, 6, 2, 9, 11.

Regarding your second issue: You pass in two edges, but 3 edge features. So that‘s an expected crash and isn‘t related to symmetric adjacencies.

[rusty1s]

Edge 0 and Edge 4 are sampled for node 0 in layer 1. The NeighborSampler will recursively sample neighbors/edges for all nodes encountered in all previous layers, e.g., for applying skip-connections.