[Graphbolt] Support loading heterogeneous attributes in sampling graph. (#6873)

python/dgl/graphbolt/impl/fused_csc_sampling_graph.py

@@ -6,7 +6,7 @@ import torch
 
 from dgl.utils import recursive_apply
 
-from ...base import EID, ETYPE
+from ...base import EID, ETYPE, NID, NTYPE
 from ...convert import to_homogeneous
 from ...heterograph import DGLGraph
 from ..base import etype_str_to_tuple, etype_tuple_to_str, ORIGINAL_EDGE_ID
@@ -1117,7 +1117,9 @@ def from_dglgraph(
 ) -> FusedCSCSamplingGraph:
     """Convert a DGLGraph to FusedCSCSamplingGraph."""
 
-    homo_g, ntype_count, _ = to_homogeneous(g, return_count=True)
+    homo_g, ntype_count, _ = to_homogeneous(
+        g, ndata=g.ndata, edata=g.edata, return_count=True
+    )
 
     if is_homogeneous:
         node_type_to_id = None
@@ -1147,8 +1149,13 @@ def from_dglgraph(
     )
 
     node_attributes = {}
-
     edge_attributes = {}
+    for feat_name, feat_data in homo_g.ndata.items():
+        if feat_name not in (NID, NTYPE):
+            node_attributes[feat_name] = feat_data
+    for feat_name, feat_data in homo_g.edata.items():
+        if feat_name not in (EID, ETYPE):
+            edge_attributes[feat_name] = feat_data
     if include_original_edge_id:
         # Assign edge attributes according to the original eids mapping.
         edge_attributes[ORIGINAL_EDGE_ID] = torch.index_select(

python/dgl/graphbolt/impl/ondisk_dataset.py

@@ -129,14 +129,41 @@ def preprocess_ondisk_dataset(
                     graph_feature["format"],
                     in_memory=in_memory,
                 )
+                if is_homogeneous:
                     g.ndata[graph_feature["name"]] = node_data
+                else:
+                    g.nodes[graph_feature["type"]].data[
+                        graph_feature["name"]
+                    ] = node_data
             if graph_feature["domain"] == "edge":
                 edge_data = read_data(
                     os.path.join(dataset_dir, graph_feature["path"]),
                     graph_feature["format"],
                     in_memory=in_memory,
                 )
+                if is_homogeneous:
                     g.edata[graph_feature["name"]] = edge_data
+                else:
+                    g.edges[etype_str_to_tuple(graph_feature["type"])].data[
+                        graph_feature["name"]
+                    ] = edge_data
+        if not is_homogeneous:
+            # For homogeneous graph, a node/edge feature must cover all
+            # node/edge types.
+            for feat_name, feat_data in g.ndata.items():
+                existing_types = set(feat_data.keys())
+                assert existing_types == set(g.ntypes), (
+                    f"Node feature {feat_name} does not cover all node types."
+                    + f"Existing types: {existing_types}."
+                    + f"Expected types: {g.ntypes}."
+                )
+            for feat_name, feat_data in g.edata.items():
+                existing_types = set(feat_data.keys())
+                assert existing_types == set(g.canonical_etypes), (
+                    f"Edge feature {feat_name} does not cover all edge types."
+                    + f"Existing types: {existing_types}."
+                    + f"Expected types: {g.etypes}."
+                )
 
     # 4. Convert the DGLGraph to a FusedCSCSamplingGraph.
     fused_csc_sampling_graph = from_dglgraph(