[Bug] Default value of `num_bases` can crash RelGraphConv module

🐛 Bug

As stated in the Doc, the optional input argument num_bases of RelGraphConv is None by default. However, if we actually set num_bases=None or simply not specify it, the code will crash…

It crashed because when num_bases=None, it will raise the value error at here: https://github.com/dmlc/dgl/blob/2cf05c53420546443a267d91ec624887d0078278/python/dgl/nn/pytorch/linear.py#L93-L94

We may need to fix the doc and RelGraphConv interface, to make it clear that num_bases is not optional (definitely not None) and has to be specified as a positive integer<=num. relations .

To Reproduce

Step(s) to reproduce the behavior:

1. whenever construct RelGraphConv as: conv = RelGraphConv(in_feat, out_feat, num_rels, num_bases=None) or conv = RelGraphConv(in_feat, out_feat, num_rels)

Output stack:

Traceback (most recent call last):
  File "entity.py", line 93, in <module>
    model = RGCN(in_size, 16, out_size, num_rels).to(device)
  File "entity.py", line 15, in __init__
    self.conv1 = RelGraphConv(h_dim, h_dim, num_rels, regularizer='basis',
  File "/opt/conda/lib/python3.8/site-packages/dgl/nn/pytorch/conv/relgraphconv.py", line 105, in __init__
    self.linear_r = TypedLinear(in_feat, out_feat, num_rels, regularizer, num_bases)
  File "/opt/conda/lib/python3.8/site-packages/dgl/nn/pytorch/linear.py", line 94, in __init__
    raise ValueError('Missing "num_bases" for basis regularization.')
ValueError: Missing "num_bases" for basis regularization.

Expected behavior

It should not crash with a default value.

Environment

• DGL Version (e.g., 1.0): 0.9
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): Pytorch 1.13
• OS (e.g., Linux): Linux
• How you installed DGL (conda, pip, source): source

[chang-l]

Right, this is what I understood. I will resolve this with a bugfix PR.

[mufeili]

As @jermainewang said, num_bases should also be optional. If it is not provided, it should be set to num_rels.