@@ -57,7 +56,6 @@ The construction function will do the following:
self._aggre_type = aggregator_type
self.norm = norm
self.activation = activation
self._allow_zero_in_degree = allow_zero_in_degree
In construction function, we first need to set the data dimensions. For
general Pytorch module, the dimensions are usually input dimension,
...
...
@@ -115,7 +113,7 @@ DGL NN Module Forward Function
In NN module, ``forward()`` function does the actual message passing and
computating. Compared with Pytorch’s NN module which usually takes
tensors as the parameters, DGL NN module takes an additional parameter
`DGLGraph <https://docs.dgl.ai/api/python/graph.html>`__. The
:class:`dgl.DGLGraph`. The
workload for ``forward()`` function can be splitted into three parts:
- Graph checking and graph type specification.
...
...
@@ -133,37 +131,16 @@ Graph checking and graph type specification
def forward(self, graph, feat):
with graph.local_scope():
# Graph checking
if not self._allow_zero_in_degree:
if (graph.in_degrees() == 0).any():
raise DGLError('There are 0-in-degree nodes in the graph,
'output for those nodes will be invalid.'
'This is harmful for some applications, '
'causing silent performance regression.'
'Adding self-loop on the input graph by calling
'`g = dgl.add_self_loop(g)` will resolve the issue.'
'Setting ``allow_zero_in_degree`` to be `True`
'when constructing this module will suppress the '
'check and let the code run.')
# Specify graph type then expand input feature according to graph type
feat_src, feat_dst = expand_as_pair(feat, graph)
**This part of code is usually shared by all the NN modules.**
``forward()`` needs to handle many corner cases on the input that can
lead to invalid values in computing and message passing. The above
example handles the case where there are 0-in-degree nodes in the input
graph.
When a node has 0-in-degree, the ``mailbox`` will be empty and the
reduce function will not produce valid values. For example, if the
reduce function is ``max``, the output for the 0-in-degree nodes
will be ``-inf``.
lead to invalid values in computing and message passing. One typical check in conv modules like :class:`~dgl.nn.pytorch.conv.GraphConv` is to verify no 0-in-degree node in the input graph. When a node has 0-in-degree, the ``mailbox`` will be empty and the reduce function will produce all-zero values. This may cause silent regression in model performance. However, in :class:`~dgl.nn.pytorch.conv.SAGEConv` module, the aggregated representation will be concatenated with the original node feature, the output of ``forward()`` will not be all-zero. No such check is needed in this case.
DGL NN module should be reusable across different types of graph input
