[Sparse] Refactor SIGN model. (#4921)

Co-authored-by: Steve <ubuntu@ip-172-31-34-29.ap-northeast-1.compute.internal>

examples/sparse/sign.py

@@ -57,24 +57,31 @@ def evaluate(g, pred):
     return val_acc, test_acc
 
 
-def train(g, model):
-    labels = g.ndata["label"]
+def train(g, X_sign, model):
+    label = g.ndata["label"]
     train_mask = g.ndata["train_mask"]
     optimizer = Adam(model.parameters(), lr=3e-3)
 
     for epoch in range(10):
+        # Switch the model to training mode.
+        model.train()
+
         # Forward.
         logits = model(X_sign)
 
         # Compute loss with nodes in training set.
-        loss = F.cross_entropy(logits[train_mask], labels[train_mask])
+        loss = F.cross_entropy(logits[train_mask], label[train_mask])
 
         # Backward.
         optimizer.zero_grad()
         loss.backward()
         optimizer.step()
 
+        # Switch the model to evaluating mode.
+        model.eval()
+
         # Compute prediction.
+        logits = model(X_sign)
         pred = logits.argmax(1)
 
         # Evaluate the prediction.
@@ -117,4 +124,4 @@ if __name__ == "__main__":
     model = SIGN(in_size, out_size, r).to(dev)
 
     # Kick off training.
-    train(g, model)
+    train(g, X_sign, model)