diff --git a/examples/mxnet/gcn/README.md b/examples/mxnet/gcn/README.md
index f6d2878199c7..c389975d5376 100644
--- a/examples/mxnet/gcn/README.md
+++ b/examples/mxnet/gcn/README.md
@@ -21,6 +21,7 @@ The folder contains three implementations of GCN:
 - `gcn_mp.py` uses user-defined message and reduce functions.
 - `gcn_spmv.py` improves from `gcn_mp.py` by using DGL's builtin functions
    so SPMV optimization could be applied.
+Modify `train.py` to switch between different implementations.
 
 The provided implementation in `gcn_concat.py` is a bit different from the
 original paper for better performance, credit to @yifeim and @ZiyueHuang.
@@ -29,7 +30,7 @@ Results
 -------
 Run with following (available dataset: "cora", "citeseer", "pubmed")
 ```bash
-DGLBACKEND=mxnet python3 train.py --dataset cora --gpu 0
+DGLBACKEND=mxnet python3 train.py --dataset cora --gpu 0 --self-loop
 ```
 
 * cora: ~0.810 (paper: 0.815)
diff --git a/examples/mxnet/gcn/train.py b/examples/mxnet/gcn/train.py
index 4eb120d82dce..dd574d9150a1 100644
--- a/examples/mxnet/gcn/train.py
+++ b/examples/mxnet/gcn/train.py
@@ -53,7 +53,8 @@ def main(args):
 
     # create GCN model
     g = DGLGraph(data.graph)
-    g.add_edges(g.nodes(), g.nodes())
+    if args.self_loop:
+        g.add_edges(g.nodes(), g.nodes())
     # normalization
     degs = g.in_degrees().astype('float32')
     norm = mx.nd.power(degs, -0.5)
@@ -120,6 +121,9 @@ def main(args):
             help="number of hidden gcn layers")
     parser.add_argument("--weight-decay", type=float, default=5e-4,
             help="Weight for L2 loss")
+    parser.add_argument("--self-loop", action='store_true',
+            help="graph self-loop (default=False)")
+    parser.set_defaults(self_loop=False)
     args = parser.parse_args()
 
     print(args)
diff --git a/examples/pytorch/gcn/README.md b/examples/pytorch/gcn/README.md
index d2f7af8459bb..ea621c066301 100644
--- a/examples/pytorch/gcn/README.md
+++ b/examples/pytorch/gcn/README.md
@@ -28,7 +28,7 @@ Results
 
 Run with following (available dataset: "cora", "citeseer", "pubmed")
 ```bash
-python train.py --dataset cora --gpu 0
+python train.py --dataset cora --gpu 0 --self-loop
 ```
 
 * cora: ~0.810 (0.79-0.83) (paper: 0.815)
diff --git a/examples/pytorch/gcn/train.py b/examples/pytorch/gcn/train.py
index c4ba300d8fba..153679db968b 100644
--- a/examples/pytorch/gcn/train.py
+++ b/examples/pytorch/gcn/train.py
@@ -57,7 +57,8 @@ def main(args):
     g = DGLGraph(data.graph)
     n_edges = g.number_of_edges()
     # add self loop
-    g.add_edges(g.nodes(), g.nodes())
+    if args.self_loop:
+        g.add_edges(g.nodes(), g.nodes())
     # normalization
     degs = g.in_degrees().float()
     norm = torch.pow(degs, -0.5)
@@ -128,6 +129,9 @@ def main(args):
             help="number of hidden gcn layers")
     parser.add_argument("--weight-decay", type=float, default=5e-4,
             help="Weight for L2 loss")
+    parser.add_argument("--self-loop", action='store_true',
+            help="graph self-loop (default=False)")
+    parser.set_defaults(self_loop=False)
     args = parser.parse_args()
     print(args)