reproduce paper results

examples/citation_benchmark/README.md

@@ -6,7 +6,21 @@ This page tries to reproduce all the **Graph Neural Network** paper for Citation
 
 
 
-All datasets are runned with public split of  **semi-supervised** settings.
+All datasets are runned with public split of  **semi-supervised** settings. And we report the averarge accuracy by running 10 times.
+
+
+
+# Experiment Results
+
+| Model                                                        | Cora         | Pubmed       | Citeseer     | Remarks                                                   |
+| ------------------------------------------------------------ | ------------ | ------------ | ------------ | --------------------------------------------------------- |
+| [Vanilla GCN (Kipf 2017)](https://openreview.net/pdf?id=SJU4ayYgl ) | 0.807(0.010) | 0.794(0.003) | 0.710(0.007) |                                                           |
+| [GAT (Veličković 2017)](https://arxiv.org/pdf/1710.10903.pdf) | 0.834(0.004) | 0.772(0.004) | 0.700(0.006) |                                                           |
+| [SGC(Wu 2019)](https://arxiv.org/pdf/1902.07153.pdf)         | 0.818(0.000) | 0.782(0.000) | 0.708(0.000) |                                                           |
+| [APPNP (Johannes 2018)](https://arxiv.org/abs/1810.05997)    | 0.846(0.003) | 0.803(0.002) | 0.719(0.003) | Almost the same with  the results reported in Appendix E. |
+| [GCNII (64 Layers, 1500 Epochs, Chen 2020)](https://arxiv.org/pdf/2007.02133.pdf) | 0.846(0.003) | 0.798(0.003) | 0.724(0.006) |                                                           |
+
+
 
 
 

examples/citation_benchmark/config/gat.yaml

@@ -6,4 +6,4 @@ feat_drop: 0.6
 attn_drop: 0.6
 num_heads: 8
 hidden_size: 8
-edge_dropout: 0.1
+edge_dropout: 0.0

examples/citation_benchmark/train.py

@@ -131,7 +131,9 @@ def main(args, config):
                   cal_test_acc[np.argmin(cal_val_loss)]))
 
         best_test.append(cal_test_acc[np.argmin(cal_val_loss)])
-    log.info("Best Test Accuracy: %f ( stddev: %f )" % (np.mean(best_test), np.std(best_test)))
+    log.info("Dataset: %s Best Test Accuracy: %f ( stddev: %f )" % (args.dataset, np.mean(best_test), np.std(best_test)))
+    print("Dataset: %s Best Test Accuracy: %f ( stddev: %f )" % (args.dataset, np.mean(best_test), np.std(best_test)))
+    
     
 
 
@@ -142,7 +144,7 @@ if __name__ == '__main__':
     parser.add_argument("--use_cuda", action='store_true', help="use_cuda")
     parser.add_argument("--conf", type=str, help="config file for models")
     parser.add_argument("--epoch", type=int, default=200, help="Epoch")
-    parser.add_argument("--runs", type=int, default=5, help="runs")
+    parser.add_argument("--runs", type=int, default=10, help="runs")
     parser.add_argument("--feature_pre_normalize", type=bool, default=True, help="pre_normalize feature")
     args = parser.parse_args()
     config = edict(yaml.load(open(args.conf), Loader=yaml.FullLoader))