Error Found when using RGCN with Node Features

Questions
#1

Hi, I am trying to use RGCN for node classification with one type of node and two types of edges. All nodes have pre-defined feature with 200 dimension. I have followed the tutorial here, but have got some error as follows.

 Traceback (most recent call last):
  File "run_classifier_rgcn.py", line 363, in <module>
    main()
  File "run_classifier_rgcn.py", line 317, in main
    train(train_dataset, model, mlb, g, feats, edge_type, edge_norm, args.batch_sz, args.num_epochs, criterion, device,
  File "run_classifier_rgcn.py", line 165, in train
    output = model(text, G, feats, edge_type, edge_norm)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/lustre03/project/6001103/xdwang/MeSH_Indexing_RGCN/model.py", line 404, in forward
    label_feature = self.rgcn(g, g_node_feature, edge_type, edge_norm)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/lustre03/project/6001103/xdwang/MeSH_Indexing_RGCN/model.py", line 364, in forward
    h = layer(g, h, r, norm)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/dgl/nn/pytorch/conv/relgraphconv.py", line 282, in forward
    g.update_all(self.message_func, fn.sum(msg='msg', out='h'))
  File "/home/xdwang/dgl/lib/python3.8/site-packages/dgl/heterograph.py", line 4499, in update_all
    ndata = core.message_passing(g, message_func, reduce_func, apply_node_func)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/dgl/core.py", line 291, in message_passing
    msgdata = invoke_edge_udf(g, ALL, g.canonical_etypes[0], mfunc, orig_eid=orig_eid)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/dgl/core.py", line 82, in invoke_edge_udf
    return func(ebatch)
  File "/home/xdwang/dgl/lib/python3.8/site-packages/dgl/nn/pytorch/conv/relgraphconv.py", line 209, in basis_message_func
    sub_msg = th.matmul(src, w)
RuntimeError: mat1 dim 1 must match mat2 dim 0

I have followed the tutorial and the only different is that I used my own dataset. I have created a heterogenous graph using dgl.heterograph(). My graph looks like Graph(num_nodes={'mesh': 29368}, num_edges={('mesh', 'cooccurrence', 'mesh'): 861, ('mesh', 'hierarchy', 'mesh'): 39920}, metagraph=[('mesh', 'mesh', 'cooccurrence'), ('mesh', 'mesh', 'hierarchy')]) and the node features look likeG.ndata['feat'].shape --->torch.Size([29368, 200])
Thanks!

#2

The example you used for reference does not use pre-defined input node features and learns node embeddings from scratch. It defines the first RGCN layer as follows

RelGraphConv(self.num_nodes, self.h_dim, self.num_rels, "basis",
             self.num_bases, activation=F.relu, self_loop=self.use_self_loop,
             dropout=self.dropout)

self.num_nodes actually specifies the input node feature size as one-hot encodings are used. Have you changed it to 200?

#3

Thanks! I passed the wrong number as I mistakenly thought num_nodes is the number of nodes in the graph at the first time.

Another question, do you happen to know what might cause this error File "dgl/_ffi/_cython/./base.pxi", line 155, in dgl._ffi._cy3.core.CALL dgl._ffi.base.DGLError: [12:08:08] /home/coulombc/wheels_builder/tmp.30875/python-3.8/dgl/src/array/cuda/utils.cu:19: Check failed: e == cudaSuccess || e == cudaErrorCudartUnloading: CUDA: invalid device ordinal
I think I have set the correct gpu id accordingly.
Thanks!

#4

@xdwang0726 are you using PyTorch 1.7 with CUDA 11?

#5

I am using torch 1.7 with cuda 10.2

>>> print(torch.__version__)
1.7.0
>>> torch.version.cuda
'10.2'
#6

@xdwang0726 could you share how you apply nodes’ features into that tutorial? It don’t use any nodes’ features into rgcn.