[Bug] Fix Reddit

Author: dddg617

examples/graphsage/reddit_sage.py

@@ -48,15 +48,15 @@ def main(args):
     test_idx = mask_to_index(graph.test_mask)
     val_idx = mask_to_index(graph.val_mask)
 
-    train_loader = NeighborSampler(edge_index=graph.edge_index.numpy(),
-                                   node_idx=tlx.convert_to_numpy(train_idx),
+    train_loader = NeighborSampler(edge_index=graph.edge_index,
+                                   node_idx=train_idx,
                                    sample_lists=[25, 10], batch_size=2048, shuffle=True, num_workers=0)
 
-    val_loader = NeighborSampler(edge_index=graph.edge_index.numpy(),
-                                 node_idx=tlx.convert_to_numpy(val_idx),
+    val_loader = NeighborSampler(edge_index=graph.edge_index,
+                                 node_idx=val_idx,
                                  sample_lists=[-1], batch_size=2048 * 2, shuffle=False, num_workers=0)
-    test_loader = NeighborSampler(edge_index=graph.edge_index.numpy(),
-                                  node_idx=tlx.convert_to_numpy(test_idx),
+    test_loader = NeighborSampler(edge_index=graph.edge_index,
+                                  node_idx=test_idx,
                                   sample_lists=[-1], batch_size=2048 * 2, shuffle=False, num_workers=0)
 
     x = tlx.convert_to_tensor(graph.x)
@@ -78,6 +78,9 @@ def main(args):
         pbar = tqdm(total=int(len(train_loader.dataset)))
         pbar.set_description(f'Epoch {epoch:02d}')
         for dst_node, n_id, adjs in train_loader:
+            print("---------------")
+            print(adjs)
+            print(type(adjs))
             net.set_train()
             # input : sampled subgraphs, sampled node's feat
             data = {"x": tlx.gather(x, n_id),

gammagl/datasets/reddit.py

@@ -52,13 +52,13 @@ def download(self):
 
     def process(self):
         data = np.load(osp.join(self.raw_dir, 'reddit_data.npz'))
-        x = np.array(data['feature'], dtype=np.float32)
-        y = np.array(data['label'], np.int32)
+        x = tlx.convert_to_tensor(data['feature'], dtype=tlx.float32)
+        y = tlx.convert_to_tensor(data['label'], dtype=tlx.int64)
         split = np.array(data['node_types'])
 
         adj = sp.load_npz(osp.join(self.raw_dir, 'reddit_graph.npz'))
 
-        edge = np.array([adj.row, adj.col], dtype=np.int64)
+        edge = tlx.convert_to_tensor([adj.row, adj.col], dtype=tlx.int64)
 
         edge, _ = coalesce(edge, None, x.shape[0], x.shape[0])
 

gammagl/models/graphsage.py

@@ -75,28 +75,20 @@ def __init__(self, in_feat, hid_feat, out_feat, drop_rate, num_layers, name=None
 
     def forward(self, x, edgeIndices):
         for l, (layer, edgeIndex) in enumerate(zip(self.convs, edgeIndices)):
-            if tlx.BACKEND == 'torch':
-                edgeIndex.to(x.device)
             target_x = tlx.gather(x, tlx.arange(0, edgeIndex.size[1]))  # Target nodes are always placed first.
             x = layer((x, target_x), edgeIndex.edge_index)
             if l != len(self.convs) - 1:
                 x = self.dropout(x)
         return x
 
     def inference(self, feat, dataloader, cur_x):
-        if tlx.BACKEND == 'torch':
-            feat = feat.to(cur_x.device)
         for l, layer in enumerate(self.convs):
             y = tlx.zeros((feat.shape[0], self.num_class if l == len(self.convs) - 1 else self.hid_feat))
-            if tlx.BACKEND == 'torch':
-                y = y.to(feat.device)
             for dst_node, n_id, adjs in dataloader:
                 if isinstance(adjs, (List, Tuple)):
                     sg = adjs[0]
                 else:
                     sg = adjs
-                if tlx.BACKEND == 'torch':
-                    sg.to(y.device)
                 h = tlx.gather(feat, n_id)
                 target_feat = tlx.gather(h, tlx.arange(0, sg.size[1]))
                 h = layer((h, target_feat), sg.edge_index)