Dgl.add_self_loop() Throws RuntimeError: Could not infer dtype of numpy.int64

[rockinbarisax]

Hey there!
I have my own dataset I created using the introductory tutorial. But when I try to use it in my model, I get the error that “There are 0-in-degree nodes in the graph, output for those nodes will be invalid.” in order to fix the problem, it says to add “g = dgl.add_self_loop(g)” to my code. However, dgl.add_self_loop causes the error “RuntimeError: Could not infer dtype of numpy.int64”, which is interesting becuase I’m not using any int’s in my graphs whatesever.

Here’s is what I get when I print the graph:
Graph(num_nodes=24, num_edges=50,
ndata_schemes={‘atomicNums’: Scheme(shape=(), dtype=torch.float64), ‘temperature’: Scheme(shape=(), dtype=torch.float64)}
edata_schemes={‘bondOrders’: Scheme(shape=(), dtype=torch.float64)})

I’m using the following versions:
torchVersion : 1.8.1 (PyTorch)
DGLVersion : 0.7.1
PythonVerision: 3.6.5 |Anaconda, Inc.| (default, Mar 29 2018, 13:32:41) [MSC v.1900 64 bit (AMD64)]

Any ideas?

[rockinbarisax]

I should add, that I’ve been getting the same " [RuntimeError: Could not infer dtype of numpy.int64]" error in a variety of different situations. For example, I get the same error when completing tutorial found at Training a GNN for Graph Classification — DGL 0.7.1 documentation . In particular, the dataset = dgl.data.GINDataset(‘PROTEINS’, self_loop=True) command throws the error.

[rockinbarisax]

UPDATE: It appears that using a different version of DGL (0.6.0post1) fixed the problem. (Also switched to the most recent LTE of PyTorch 1.8.2, as well as Python 3.9.7)

[VoVAllen]

Sorry this seems a bug for DGL on windows. Could you raise an issue in our github repo? Thanks