MultiLayerNeighborSampler usage in heterogenous graph

I want to use the MultiLayerNeighborSampler on a heterogenous graph.
In the first layer, I’d like to sample x edges irrespective of edge type(i.e. total x incoming edges for every node). I have no preference per edge type.
How do I do that?

Will using dgl.dataloading.MultiLayerNeighborSampler([x, y]) give me the correct result?
Or will this sample x edges per edge type for each node?

Edit to add -
The source says

If a single int is given, DGL will sample this number of edges for each node for every edge type.

How do I sample x edges per node?

1 Like

You currently cannot do that with MultiLayerNeighborSampler on a heterogeneous graph. You can consider homogenizing your graph via to_homogeneous, which preserves node and edge type information as integral node/edge features.

Would it be possible to write a custom sampler for heterogenous graphs? How do I go about writing it?

Yes. You can take a look here for an example of custom neighbor sampler.

I would imagine that an efficient edge-type-irrelevant sampler will still involve keeping a homogenized version of the original heterogeneous graph. My intuition is that you only need to implement the sample_frontier method: after you perform (homogeneous) neighbor sampling on the homogenized graph, you can then heterogenize it back with dgl.to_heterogeneous.

Hi, this is my attempt at the edge-type-irrelevant neighbor sampler.

    class MultiLayerNeighborSampler(dgl.dataloading.BlockSampler):
        def __init__(self, fanouts, g):
            super().__init__(len(fanouts), return_eids=False)

            self.fanouts = fanouts
            self.homo_g = dgl.to_homogeneous(g)
            self.mapping = self.get_mapping() # Map heterogenous node ids to homogenous node ids using _ID and _TYPE
        
        def get_mapping(self):
            ...
            return mapping

        def sample_frontier(self, block_id, g, seed_nodes):
            '''
            seed_nodes are in the heterogenous format 
            We convert them to homogenous and then sample neighbors
            and convert the frontier back to hetero
            '''
            seed_nodes = ... # convert to homogeneous ids using self.mapping

            fanout = self.fanouts[block_id]
            if fanout is None:
                frontier = dgl.in_subgraph(self.homo_g, seed_nodes)
            else:
                frontier = dgl.sampling.sample_neighbors(self.homo_g, seed_nodes, fanout)

            # Convert back to hetero
            frontier = dgl.to_heterogeneous(frontier, g.ntypes, g.etypes)
            return frontier

I get the following error during the dgl.sampling.sample_neighbors() call -

Traceback (most recent call last):
  File "hetero.py", line 660, in <module>
    train(args, devices)
  File "hetero.py", line 525, in train
    run(0, n_gpus, args, devices, data)
  File "hetero.py", line 412, in run
    for step, (input_nodes, pos_graph, neg_graph, blocks) in enumerate(train_dataloader):
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/dataloading/pytorch/__init__.py", line 173, in __next__
    result_ = next(self.iter_)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
    data = self._next_data()
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 1085, in _next_data
    return self._process_data(data)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 1111, in _process_data
    data.reraise()
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/_utils.py", line 428, in reraise
    raise self.exc_type(msg)
dgl._ffi.base.DGLError: Caught DGLError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 198, in _worker_loop
    data = fetcher.fetch(index)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 47, in fetch
    return self.collate_fn(data)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/dataloading/pytorch/__init__.py", line 136, in collate
    result = super().collate(items)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/dataloading/dataloader.py", line 720, in collate
    return self._collate_with_negative_sampling(items)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/dataloading/dataloader.py", line 676, in _collate_with_negative_sampling
    self.g_sampling, seed_nodes, exclude_eids=exclude_eids)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/dataloading/dataloader.py", line 253, in sample_blocks
    block = transform.to_block(frontier, seed_nodes)
  File "/home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/transform.py", line 1829, in to_block
    g._graph, dst_node_ids_nd, include_dst_in_src)
  File "dgl/_ffi/_cython/./function.pxi", line 287, in dgl._ffi._cy3.core.FunctionBase.__call__
  File "dgl/_ffi/_cython/./function.pxi", line 222, in dgl._ffi._cy3.core.FuncCall
  File "dgl/_ffi/_cython/./function.pxi", line 211, in dgl._ffi._cy3.core.FuncCall3
  File "dgl/_ffi/_cython/./base.pxi", line 155, in dgl._ffi._cy3.core.CALL
dgl._ffi.base.DGLError: [08:24:55] /opt/dgl/src/graph/transform/to_bipartite.cc:95: Check failed: new_dst.Ptr<IdType>()[i] != -1 (-1 vs. -1) : Node 12547 does not exist in `rhs_nodes`. Argument `rhs_nodes` must contain all the edge destination nodes.
Stack trace:
  [bt] (0) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x4f) [0x7f3682f3c0ff]
  [bt] (1) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(+0xd4ecc0) [0x7f368373fcc0]
  [bt] (2) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(std::tuple<std::shared_ptr<dgl::BaseHeteroGraph>, std::vector<dgl::runtime::NDArray, std::allocator<dgl::runtime::NDArray> >, std::vector<dgl::runtime::NDArray, std::allocator<dgl::runtime::NDArray> > > dgl::transform::ToBlock<(DLDeviceType)1, long>(std::shared_ptr<dgl::BaseHeteroGraph>, std::vector<dgl::runtime::NDArray, std::allocator<dgl::runtime::NDArray> > const&, bool)+0x3a) [0x7f36837409ba]
  [bt] (3) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(+0xd50281) [0x7f3683741281]
  [bt] (4) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(+0xd50ce4) [0x7f3683741ce4]
  [bt] (5) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/libdgl.so(DGLFuncCall+0x48) [0x7f3683628a98]
  [bt] (6) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/_ffi/_cy3/core.cpython-36m-x86_64-linux-gnu.so(+0x16873) [0x7f367f8ff873]
  [bt] (7) /home/ubuntu/anaconda3/envs/new/lib/python3.6/site-packages/dgl/_ffi/_cy3/core.cpython-36m-x86_64-linux-gnu.so(+0x16b9b) [0x7f367f8ffb9b]
  [bt] (8) python(_PyObject_FastCallDict+0x8b) [0x55c97f46e0bb]

This happens only if I change the Neighbor sample to the one above. This is the only change to an otherwise heterogeneous GNN implementation. It works normally if I use any of dgls samplers.

Could you help me figure this one out? Thanks a lot

Figured out the problem. The reason is that to_heterogeneous(to_homogeneous(hg)) actually reorders the nodes, so essentially to_heterogeneous(sample_neighbors(to_homogeneous(hg))) no longer becomes a subgraph of the original hg.

I’m fixing it in [Feature] Make to_heterogeneous(to_homogeneous(hg)) return hg by BarclayII · Pull Request #2958 · dmlc/dgl · GitHub.

1 Like

Ah, I see. Thank you for fixing it so promptly @BarclayII !

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.