How to apply NN modules on DGLHeteroGraph?

Questions
#1

I have a DGLHeteroGraph, and I want to process it as a homo-graph after some operation and calling dgl.to_homo(), however this method still returns a DGLHeteroGraph. If I want to use NN modules on it,

homo_g = dgl.to_homo(g)
self.gat1 = GATConv(in_feats=self.entity_dim, out_feats=8, num_heads=8, activation=nn.ELU(),
                            feat_drop=0.6, attn_drop=0.6)
node_feats = self.gat1(homo_g, homo_g.nodes['_N'].data['feature'])

the following error would occur,

File "/home/weiguoao/anaconda3/envs/wga/lib/python3.7/site-packages/torch/nn/modules/module.py", line 532, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/weiguoao/projects/KG-GAT/models.py", line 167, in forward
    node_feats = self.gat1(homo_g, homo_g.nodes['_N'].data['feature'])
  File "/home/weiguoao/anaconda3/envs/wga/lib/python3.7/site-packages/torch/nn/modules/module.py", line 532, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/weiguoao/anaconda3/envs/wga/lib/python3.7/site-packages/dgl/nn/pytorch/conv/gatconv.py", line 111, in forward
    graph.edata['a'] = self.attn_drop(edge_softmax(graph, e))
  File "/home/weiguoao/anaconda3/envs/wga/lib/python3.7/site-packages/dgl/nn/pytorch/softmax.py", line 179, in edge_softmax
    return EdgeSoftmax.apply(graph, logits, eids)
  File "/home/weiguoao/anaconda3/envs/wga/lib/python3.7/site-packages/dgl/nn/pytorch/softmax.py", line 52, in forward
    gidx = g._graph.get_immutable_gidx(utils.to_dgl_context(score.device))
  File "dgl/_ffi/_cython/./object.pxi", line 64, in dgl._ffi._cy3.core.ObjectBase.__getattr__
AttributeError: '<class 'dgl.heterograph_index.HeteroGraphIndex'>' object has no attribute 'get_immutable_gidx'

is there any way I could apply NN modules on homo DGLHeteroGraph? Maybe I can convert DGLHeteroGraph to DGLGraph?

OS: Ubuntu 16.04
DGL version: dgl-cu100 0.4.2

#2

Hi, currently our SAGEConv model supports DGLHeteroGraph.
For our models, I’m working on refactoring the code to let them support heterographs, please stay tuned.

#3

Hi,

The problem should be fixed in the nightly builds. Could you try installing the nightly build and run it again?

pip install --pre dgl  (or pip install --pre dgl-cu100 for CUDA 10.0 etc.)
#4

My previously DGLHeteroGraph was saved as .pkl. I installed the nightly builds and tried to load from .pkl, this error occured,

  File "/home/weiguoao/projects/KG-GAT/dataset.py", line 33, in __init__
    self.g = pickle.load(f)
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/heterograph_index.py", line 32, in __setstate__
    self.__init_handle_by_constructor__(_CAPI_DGLHeteroUnpickle, state)
  File "dgl/_ffi/_cython/./object.pxi", line 86, in dgl._ffi._cy3.core.ObjectBase.__init_handle_by_constructor__
  File "dgl/_ffi/_cython/./function.pxi", line 244, in dgl._ffi._cy3.core.ConstructorCall
  File "dgl/_ffi/_cython/./function.pxi", line 222, in dgl._ffi._cy3.core.FuncCall
  File "dgl/_ffi/_cython/./function.pxi", line 210, in dgl._ffi._cy3.core.FuncCall3
  File "dgl/_ffi/_cython/./function.pxi", line 135, in dgl._ffi._cy3.core.make_arg
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in convert_to_object
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in <listcomp>
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in convert_to_object
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in <listcomp>
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in convert_to_object
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 38, in <listcomp>
    value = [convert_to_object(x) for x in value]
  File "/home/weiguoao/anaconda3/envs/dgl-pre/lib/python3.7/site-packages/dgl/_ffi/object_generic.py", line 51, in convert_to_object
    return _api_internal._Value(value)
  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 210, in dgl._ffi._cy3.core.FuncCall3
  File "dgl/_ffi/_cython/./function.pxi", line 154, in dgl._ffi._cy3.core.make_arg
TypeError: Don't know how to handle type <class 'dgl.utils.Index'>

I rebuild the graph and try to save&load, the same error still occured.

#5

Hi,

I don’t think we maintain the compatibility of pickle dumps across DGL versions.

However, since you mentioned rebuilding the graph and saving/loading afterwards, I tried saving and loading with a heterograph with pickle and could not reproduce the error. Could you tell us how you save and load the graph? Thanks!

#6

I found that the problem occurred because I save the graph using 0.4.2 but load with 0.4.3, thank you for your answer anyway.
By the way, is there an official API for saving&loading heterograph?