[Bugs] Fix distributed example error and import error (#3783)

* fix * raise an error * fix docserver crash
dmlc · Feb 28, 2022 · 62e23bd · 62e23bd
1 parent 7ab1034
commit 62e23bd
Show file tree

Hide file tree

Showing 5 changed files with 28 additions and 26 deletions.
diff --git a/include/dgl/aten/array_ops.h b/include/dgl/aten/array_ops.h
@@ -356,25 +356,20 @@ IdArray VecToIdArray(const std::vector<T>& vec,
 }
 
 /*!
- * \brief Get the context of the first non-null array, and check if the non-null arrays'
+ * \brief Get the context of the first array, and check if the non-null arrays'
  * contexts are the same.
- *
- * Throws an error if all the arrays are null arrays.
  */
 inline DLContext GetContextOf(const std::vector<IdArray>& arrays) {
   bool first = true;
   DLContext result;
   for (auto& array : arrays) {
-    if (IsNullArray(array))
-      continue;
     if (first) {
       first = false;
       result = array->ctx;
     } else {
       CHECK_EQ(array->ctx, result) << "Context of the input arrays are different";
     }
   }
-  CHECK(!first) << "All input arrays are empty.";
   return result;
 }
 

diff --git a/python/dgl/__init__.py b/python/dgl/__init__.py
@@ -49,7 +49,7 @@
 from .data.utils import save_graphs, load_graphs
 from . import optim
 from .frame import LazyFeature
-from .utils import recursive_apply
+from .utils import apply_each
 
 from ._deprecate.graph import DGLGraph as DGLGraphStale
 from ._deprecate.nodeflow import *
diff --git a/python/dgl/sampling/neighbor.py b/python/dgl/sampling/neighbor.py
@@ -309,12 +309,17 @@ def _sample_neighbors(g, nodes, fanout, edge_dir='in', prob=None, replace=False,
         nodes = {g.ntypes[0] : nodes}
 
     nodes = utils.prepare_tensor_dict(g, nodes, 'nodes')
+    if len(nodes) == 0:
+        raise ValueError(
+            "Got an empty dictionary in the nodes argument. "
+            "Please pass in a dictionary with empty tensors as values instead.")
+    ctx = utils.to_dgl_context(F.context(next(iter(nodes.values()))))
     nodes_all_types = []
     for ntype in g.ntypes:
         if ntype in nodes:
             nodes_all_types.append(F.to_dgl_nd(nodes[ntype]))
         else:
-            nodes_all_types.append(nd.array([], ctx=nd.cpu()))
+            nodes_all_types.append(nd.array([], ctx=ctx))
 
     if isinstance(fanout, nd.NDArray):
         fanout_array = fanout
@@ -354,7 +359,7 @@ def _sample_neighbors(g, nodes, fanout, edge_dir='in', prob=None, replace=False,
             if etype in exclude_edges:
                 excluded_edges_all_t.append(F.to_dgl_nd(exclude_edges[etype]))
             else:
-                excluded_edges_all_t.append(nd.array([], ctx=nd.cpu()))
+                excluded_edges_all_t.append(nd.array([], ctx=ctx))
 
     subgidx = _CAPI_DGLSampleNeighbors(g._graph, nodes_all_types, fanout_array,
                                        edge_dir, prob_arrays, excluded_edges_all_t, replace)

diff --git a/tutorials/multi/1_graph_classification.py b/tutorials/multi/1_graph_classification.py
@@ -227,17 +227,18 @@ def main(rank, world_size, dataset, seed=0):
 
 ###############################################################################
 # Finally we load the dataset and launch the processes.
-# 
-
-if __name__ == '__main__':
-    import torch.multiprocessing as mp
-
-    from dgl.data import GINDataset
-
-    num_gpus = 4
-    procs = []
-    dataset = GINDataset(name='IMDBBINARY', self_loop=False)
-    mp.spawn(main, args=(num_gpus, dataset), nprocs=num_gpus)
+#
+# .. code:: python
+#
+#    if __name__ == '__main__':
+#        import torch.multiprocessing as mp
+#    
+#        from dgl.data import GINDataset
+#    
+#        num_gpus = 4
+#        procs = []
+#        dataset = GINDataset(name='IMDBBINARY', self_loop=False)
+#        mp.spawn(main, args=(num_gpus, dataset), nprocs=num_gpus)
 
 # Thumbnail credits: DGL
 # sphinx_gallery_thumbnail_path = '_static/blitz_5_graph_classification.png'
diff --git a/tutorials/multi/2_node_classification.py b/tutorials/multi/2_node_classification.py
@@ -245,12 +245,13 @@ def run(proc_id, devices):
 #    Python’s built-in ``multiprocessing`` except that it handles the
 #    subtleties between forking and multithreading in Python.
 # 
-
-# Say you have four GPUs.
-if __name__ == '__main__':
-    num_gpus = 4
-    import torch.multiprocessing as mp
-    mp.spawn(run, args=(list(range(num_gpus)),), nprocs=num_gpus)
+# .. code:: python
+#
+#    # Say you have four GPUs.
+#    if __name__ == '__main__':
+#        num_gpus = 4
+#        import torch.multiprocessing as mp
+#        mp.spawn(run, args=(list(range(num_gpus)),), nprocs=num_gpus)
 
 # Thumbnail credits: Stanford CS224W Notes
 # sphinx_gallery_thumbnail_path = '_static/blitz_1_introduction.png'