NeighborLoader returns all empty except seed nodes

[cctony1]

I'm able to use the neighbor loader with the ogb mag dataset in the example here: https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/to_hetero_mag.py

However, I have a problem for my own dataset.

I created this, train_data:

HeteroData(
p={
x=[1517, 96],
y=[1517]
},
mc={ x=[102354, 14] },
rx={ x=[46659, 6] },
po={ x=[2350, 36] },
dx={ x=[3496, 22] },
d={ x=[1993, 5] },
pa={ x=[53022, 81] },
(p, WP, rx)={ edge_index=[2, 46659] },
(p, H, mc)={ edge_index=[2, 102354] },
(rx, F, d)={ edge_index=[2, 46603] },
(rx, PB, pa)={ edge_index=[2, 46659] },
(mc, W, po)={ edge_index=[2, 60928] },
(mc, W, dx)={ edge_index=[2, 41426] },
(mc, SB, pa)={ edge_index=[2, 102354] }
)

But I'm having trouble getting this to work:
train_input_nodes = ('Person', torch.arange(3)) # just 3 nodes for checking code
kwargs = {'batch_size': 1, 'num_workers': 2, 'persistent_workers': True} # batch size of 1 for checking code
train_loader = NeighborLoader(train_data, num_neighbors=[2,2], shuffle=True, input_nodes=train_input_nodes, **kwargs)

The seed nodes are sampled correctly, but the other nodes are missing, as are edge indices. Here is the data from one batch:

HeteroData(
p={
x=[1, 96],
y=[1],
batch_size=1
},
mc={ x=[0, 14] },
rx={ x=[0, 6] },
po={ x=[0, 36] },
dx={ x=[0, 22] },
d={ x=[0, 5] },
pa={ x=[0, 81] },
(p, WP, rx)={ edge_index=[2, 0] },
(p, H, mc)={ edge_index=[2, 0] },
(rx, F, d)={ edge_index=[2, 0] },
(rx, PB, pa)={ edge_index=[2, 0] },
(mc, W, po)={ edge_index=[2, 0] },
(mc, W, dx)={ edge_index=[2, 0] },
(mc, SB, pa)={ edge_index=[2, 0] }
)

Is there any (maybe obvious?) reason why the other nodes and the edges aren't populated when the neighborhood is sampled? Is it necessarily my dataset, or is there an error in how I set up the loader?

Windows 10 Enterprise
torch 1.12.1+cu113
torch-cluster 1.6.0
torch-geometric 2.0.4
torch-scatter 2.0.9
torch-sparse 0.6.15
torch-spline-conv 1.2.1
torchaudio 0.12.1+cu113
torchinfo 1.7.2
torchvision 0.13.1+cu113

[EdisonLeeeee]

train_input_nodes = ('Person', torch.arange(3)) # just 3 nodes for checking code

Didn't see this type of node in your graph.

[cctony1]

It looks like the error goes away if I remove the test(val_loader) and test(test_loader) lines.

I realize my val/test sets have a few nodes with slightly different feature dimensions from train (probably an oversight when I onehot encoded some variables in val/test that were missing categories that existed in train...)

[cctony1]

I fixed the variable issue. Training+Eval is working now. But I get validation and test accuracies of 100% (haha, wow)

I guess I'll look into data leakage. Otherwise, do you see any issue with my GNN?

[EdisonLeeeee]

No. Your GNN model looks 100% correct to me.

[cctony1]

Thanks! It's working now like a normal model (found error that led to perfect accuracy...)

[cctony1]

Since we're going from Destination to Source...

when I specify num_neighbors=[10,20] for the NeighborLoader, does the 10 or 20 refer to the seed nodes' nearest neighbors?

I had assumed 10 was nearest neighbors to seed, and 20 was the neighbors' neighbors. But since I had the message passing direction reversed in my head, I want to verify this detail.