update

README.MD

@@ -1,4 +1,6 @@
-# Graph Convolution Network for TF2
+# Graph Convolution Network for PyTorch
+
+Pytorch 1.0 with python 3.7.
 
 GCN implementation for paper: [Semi-Supervised Classification with Graph Convolutional Networks](https://arxiv.org/pdf/1609.02907.pdf)
 
@@ -7,7 +9,12 @@ GCN implementation for paper: [Semi-Supervised Classification with Graph Convolu
 | dataset       | Citeseea | Cora | Pubmed | NELL |
 |---------------|----------|------|--------|------|
 | GCN(official) | 70.3     | 81.5 | 79.0   | 66.0 |
-| This repo.    |          | 81.8 |   78.9  |      |
+| This repo.    |          | 81.7 |   78.7  |      |
+
+Hints:
+To achieve 87.7% on `Cora` dataset, pytorch need 400 epochs.
+and it train relatively slower on `Pubmed` dataset.
+
 
 # HOWTO
 ```

config.py

@@ -3,13 +3,13 @@
 args = argparse.ArgumentParser()
 args.add_argument('--dataset', default='cora')
 args.add_argument('--model', default='gcn')
-args.add_argument('--learning_rate', default=0.01)
-args.add_argument('--epochs', default=200)
-args.add_argument('--hidden1', default=16)
-args.add_argument('--dropout', default=0.5)
-args.add_argument('--weight_decay', default=5e-4)
-args.add_argument('--early_stopping', default=10)
-args.add_argument('--max_degree', default=3)
+args.add_argument('--learning_rate', type=float, default=0.01)
+args.add_argument('--epochs', type=int, default=200)
+args.add_argument('--hidden', type=int, default=16)
+args.add_argument('--dropout', type=float, default=0.5)
+args.add_argument('--weight_decay', type=float, default=5e-4)
+args.add_argument('--early_stopping', type=int, default=10)
+args.add_argument('--max_degree', type=int, default=3)
 
 
 args = args.parse_args()

model.py

@@ -3,6 +3,7 @@
 from    torch.nn import functional as F
 from    layer import GraphConvolution
 
+from    config import args
 
 class GCN(nn.Module):
 
@@ -18,14 +19,14 @@ def __init__(self, input_dim, output_dim, num_features_nonzero):
         print('num_features_nonzero:', num_features_nonzero)
 
 
-        self.layers = nn.Sequential(GraphConvolution(self.input_dim, 16, num_features_nonzero,
+        self.layers = nn.Sequential(GraphConvolution(self.input_dim, args.hidden, num_features_nonzero,
                                                      activation=F.relu,
-                                                     dropout=0.5,
+                                                     dropout=args.dropout,
                                                      is_sparse_inputs=True),
 
-                                    GraphConvolution(16, 7, num_features_nonzero,
+                                    GraphConvolution(args.hidden, output_dim, num_features_nonzero,
                                                      activation=F.relu,
-                                                     dropout=0.5,
+                                                     dropout=args.dropout,
                                                      is_sparse_inputs=False),
 
                                     )

train.py

@@ -22,17 +22,13 @@
 print('y:', y_train.shape, y_val.shape, y_test.shape)
 print('mask:', train_mask.shape, val_mask.shape, test_mask.shape)
 
-
-
 # D^-1@X
 features = preprocess_features(features) # [49216, 2], [49216], [2708, 1433]
-print('features coordinates::', features[0].shape)
-print('features data::', features[1].shape)
-print('features shape::', features[2])
 supports = preprocess_adj(adj)
 
 device = torch.device('cuda')
 train_label = torch.from_numpy(y_train).long().to(device)
+num_classes = train_label.shape[1]
 train_label = train_label.argmax(dim=1)
 train_mask = torch.from_numpy(train_mask.astype(np.int)).to(device)
 val_label = torch.from_numpy(y_val).long().to(device)
@@ -50,14 +46,15 @@
 v = torch.from_numpy(supports[1]).to(device)
 support = torch.sparse.FloatTensor(i.t(), v, supports[2]).float().to(device)
 
-print(feature)
-print(support)
+print('x :', feature)
+print('sp:', support)
 num_features_nonzero = feature._nnz()
-dropout = args.dropout
+feat_dim = feature.shape[1]
+
 
-net = GCN(1433, 7, num_features_nonzero)
+net = GCN(feat_dim, num_classes, num_features_nonzero)
 net.to(device)
-optimizer = optim.Adam(net.parameters(), lr=1e-2)
+optimizer = optim.Adam(net.parameters(), lr=args.learning_rate)
 
 net.train()
 for epoch in range(args.epochs):