I wonder how to use NeighborLoader correctly?

[PolarisRisingWar]

I haven't learnt much about older version of NeighborSampler, just directly learnt NeighborLoader which is recommended.
And I find that each batch is a PyG.data.Data format graph, which seems that I can directly put one batch in a normal model which take x,edge_index as input. So is that the correct method to apply NeighborLoader?
Besides, is factors of num_neighbors paramater means how many neighbors to sample in this hop? And all sampled nodes and sampled edges consist of the graph of the output batch? And the directed input parameter decides whether the batch data is a directed graph, the transform input parameter decides if the batch data will be transformed?
I wonder what the replace input parameter decides, if it's false, it won't sample the same neighbor twice or more? So if it's true, will the batch data has edge weight?

And I've seen the example code for NeighborSampler in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/reddit.py. It seems that the output batch was as many hops as many bipartite graphs sampled. And the code seems more complex, in training time it should use bipartite graphs to train mini-batches, in testing time it should collect node embeddings in all mini-batches for all layers, am I correct? I haven't understood it totally.

I'm changing a code in older version used NeighborSampler to a newer version. If my understanding is correct, I think it's much easier for me to use NeighborLoader than NeighborSampler. I can use it just like implementing full-batch GNN.

[rusty1s]

Besides, is factors of num_neighbors paramater means how many neighbors to sample in this hop? And all sampled nodes and sampled edges consist of the graph of the output batch? And the directed input parameter decides whether the batch data is a directed graph, the transform input parameter decides if the batch data will be transformed? I wonder what the replace input parameter decides, if it's false, it won't sample the same neighbor twice or more? So if it's true, will the batch data has edge weight?

Yes, all your understandings are correct. replace=False ensures that we sample without replacement.

And I've seen the example code for NeighborSampler in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/reddit.py. It seems that the output batch was as many hops as many bipartite graphs sampled. And the code seems more complex, in training time it should use bipartite graphs to train mini-batches, in testing time it should collect node embeddings in all mini-batches for all layers, am I correct? I haven't understood it totally.

We deprecated NeighborSampler as the underlying implementation is indeed a little bit complex, in particular, because your model code now depends on the sampler you use. The NeighborLoader changes that, and can thus be seamlessly swapped out for other loaders as well, ClusterLoader or GraphSAINTSampler.

[PolarisRisingWar]

Thank you for you answers! And I have another problem, I wonder how could I map node in output graph of NeighborLoader to node in original graph? Is it sorted as original graph if I use shuffle=False?
I'm trying to use train_mask, val_mask and test_mask when set input_nodes=None for all nodes loader when testing. And this function may also be helpful when I want to study a specific case.

[rusty1s]

You should be able to access global indices if you add an additional property to your data, e.g.:

data.n_id = torch.arange(data.num_nodes)

loader = NeighborLoader(data, ...)

for batch in loader:
    print(batch.n_id)