[BugFix] fix #3429 and update results of caregnn (#3441)

* squeeze node labels in FraudDataset

* fix RLModule

* update results in README.md

* fix KeyError in full graph training
Co-authored-by: Jinjing Zhou <VoVAllen@users.noreply.github.com>

examples/pytorch/caregnn/README.md

@@ -12,7 +12,7 @@ Dependencies
 ----------------------
 - Python 3.7.10
 - PyTorch 1.8.1
-- dgl 0.7.0
+- dgl 0.7.1
 - scikit-learn 0.23.2
 
 Dataset
@@ -48,9 +48,9 @@ The datasets used for node classification are DGL's built-in FraudDataset. The s
 
 How to run
 --------------------------------
-To run the full graph version, in the care-gnn folder, run
+To run the full graph version and use early stopping, in the care-gnn folder, run
 ```
-python main.py
+python main.py --early-stop
 ```
 
 If want to use a GPU, run
@@ -70,50 +70,57 @@ python main_sampling.py
 
 Performance
 -------------------------
-The result reported by the paper is the best validation results within 30 epochs, while ours are testing results after the max epoch specified in the table. Early stopping with patience value of 100 is applied.
+The result reported by the paper is the best validation results within 30 epochs, and the table below reports the val and test results (same setting in the paper except for the random seed, here `seed=717`). 
 
 <table>
+<thead>
   <tr>
     <th colspan="2">Dataset</th>
     <th>Amazon</th>
     <th>Yelp</th>
-	</tr >
+  </tr>
+</thead>
+<tbody>
   <tr>
-        <td>Metric</td>
+    <td>Metric (val / test)</td>
     <td>Max Epoch</td>
-	    <td>30 / 1000</td>
-	    <td>30 / 1000</td>
+    <td>30</td>
+    <td>30 </td>
   </tr>
-	<tr >
-	    <td rowspan="3">AUC</td>
+  <tr>
+    <td rowspan="3">AUC (val/test)</td>
     <td>paper reported</td>
-	    <td>89.73 / -</td>
-        <td>75.70 / -</td>
+    <td>0.8973 / -</td>
+    <td>0.7570 / -</td>
   </tr>
   <tr>
     <td>DGL full graph</td>
-	    <td>89.50 / 92.35</td>
-	    <td>69.16 / 79.91</td>
+    <td>0.8849 / 0.8922</td>
+    <td>0.6856 / 0.6867</td>
   </tr>
   <tr>
     <td>DGL sampling</td>
-	    <td>93.27 / 92.94</td>
-        <td>79.38 / 80.53</td>
+    <td>0.9350 / 0.9331</td>
+    <td>0.7857 / 0.7890</td>
   </tr>
-	<tr >
-	    <td rowspan="3">Recall</td>
+  <tr>
+    <td rowspan="3">Recall (val/test)</td>
     <td>paper reported</td>
-	    <td>88.48 / -</td>
-        <td>71.92 / -</td>
+    <td>0.8848 / -</td>
+    <td>0.7192 / -</td>
   </tr>
   <tr>
     <td>DGL full graph</td>
-	    <td>85.54 / 84.47</td>
-	    <td>69.91 / 73.47</td>
+    <td>0.8615 / 0.8544</td>
+    <td>0.6667/ 0.6619</td>
   </tr>
   <tr>
     <td>DGL sampling</td>
-	    <td>85.83 / 87.46</td>
-        <td>77.26 / 64.34</td>
+    <td>0.9130 / 0.9045</td>
+    <td>0.7537 / 0.7540</td>
   </tr>
+</tbody>
 </table>
+
+
+

examples/pytorch/caregnn/main.py

@@ -3,9 +3,10 @@ import argparse
 import torch as th
 from model import CAREGNN
 import torch.optim as optim
-from utils import EarlyStopping
 from sklearn.metrics import recall_score, roc_auc_score
 
+from utils import EarlyStopping
+
 
 def main(args):
     # Step 1: Prepare graph data and retrieve train/validation/test index ============================= #
@@ -21,10 +22,10 @@ def main(args):
         device = 'cpu'
 
     # retrieve labels of ground truth
-    labels = graph.ndata['label'].to(device).squeeze().long()
+    labels = graph.ndata['label'].to(device)
 
     # Extract node features
-    feat = graph.ndata['feature'].to(device).float()
+    feat = graph.ndata['feature'].to(device)
 
     # retrieve masks for train/validation/test
     train_mask = graph.ndata['train_mask']
@@ -121,7 +122,7 @@ if __name__ == '__main__':
     parser.add_argument("--weight_decay", type=float, default=0.001, help="Weight decay. Default: 0.001")
     parser.add_argument("--step_size", type=float, default=0.02, help="RL action step size (lambda 2). Default: 0.02")
     parser.add_argument("--sim_weight", type=float, default=2, help="Similarity loss weight (lambda 1). Default: 2")
-    parser.add_argument('--early-stop', action='store_true', default=True, help="indicates whether to use early stop")
+    parser.add_argument('--early-stop', action='store_true', default=False, help="indicates whether to use early stop")
 
     args = parser.parse_args()
     print(args)

examples/pytorch/caregnn/main_sampling.py

@@ -2,9 +2,10 @@ import dgl
 import argparse
 import torch as th
 import torch.optim as optim
+from sklearn.metrics import roc_auc_score, recall_score
+
 from utils import EarlyStopping
 from model_sampling import CAREGNN, CARESampler, _l1_dist
-from sklearn.metrics import roc_auc_score, recall_score
 
 
 def evaluate(model, loss_fn, dataloader, device='cpu'):
@@ -14,8 +15,8 @@ def evaluate(model, loss_fn, dataloader, device='cpu'):
     num_blocks = 0
     for input_nodes, output_nodes, blocks in dataloader:
         blocks = [b.to(device) for b in blocks]
-        feature = blocks[0].srcdata['feature'].float()
-        label = blocks[-1].dstdata['label'].squeeze().long()
+        feature = blocks[0].srcdata['feature']
+        label = blocks[-1].dstdata['label']
         logits_gnn, logits_sim = model(blocks, feature)
 
         # compute loss
@@ -42,10 +43,10 @@ def main(args):
         device = 'cpu'
 
     # retrieve labels of ground truth
-    labels = graph.ndata['label'].to(device).bool()
+    labels = graph.ndata['label'].to(device)
 
     # Extract node features
-    feat = graph.ndata['feature'].to(device).float()
+    feat = graph.ndata['feature'].to(device)
     layers_feat = feat.expand(args.num_layers, -1, -1)
 
     # retrieve masks for train/validation/test
@@ -58,7 +59,7 @@ def main(args):
     test_idx = th.nonzero(test_mask, as_tuple=False).squeeze(1).to(device)
 
     # Reinforcement learning module only for positive training nodes
-    rl_idx = th.nonzero(train_mask.to(device) & labels, as_tuple=False).squeeze(1)
+    rl_idx = th.nonzero(train_mask.to(device) & labels.bool(), as_tuple=False).squeeze(1)
 
     graph = graph.to(device)
 
@@ -112,8 +113,8 @@ def main(args):
 
         for input_nodes, output_nodes, blocks in train_dataloader:
             blocks = [b.to(device) for b in blocks]
-            train_feature = blocks[0].srcdata['feature'].float()
-            train_label = blocks[-1].dstdata['label'].squeeze().long()
+            train_feature = blocks[0].srcdata['feature']
+            train_label = blocks[-1].dstdata['label']
             logits_gnn, logits_sim = model(blocks, train_feature)
 
             # compute loss
@@ -184,8 +185,9 @@ if __name__ == '__main__':
     parser.add_argument("--step_size", type=float, default=0.02, help="RL action step size (lambda 2). Default: 0.02")
     parser.add_argument("--sim_weight", type=float, default=2, help="Similarity loss weight (lambda 1). Default: 0.001")
     parser.add_argument("--num_workers", type=int, default=4, help="Number of node dataloader")
-    parser.add_argument('--early-stop', action='store_true', default=True, help="indicates whether to use early stop")
+    parser.add_argument('--early-stop', action='store_true', default=False, help="indicates whether to use early stop")
 
     args = parser.parse_args()
+    th.manual_seed(717)
     print(args)
     main(args)

examples/pytorch/caregnn/model.py

@@ -43,9 +43,12 @@ class CAREConv(nn.Module):
         neigh_list = []
         for node in g.nodes():
             edges = g.in_edges(node, form='eid')
-            num_neigh = int(g.in_degrees(node) * p)
+            num_neigh = th.ceil(g.in_degrees(node) * p).int().item()
             neigh_dist = dist[edges]
+            if neigh_dist.shape[0] > num_neigh:
                 neigh_index = np.argpartition(neigh_dist.cpu().detach(), num_neigh)[:num_neigh]
+            else:
+                neigh_index = np.arange(num_neigh)
             neigh_list.append(edges[neigh_index])
         return th.cat(neigh_list)
 

examples/pytorch/caregnn/model_sampling.py

@@ -25,9 +25,12 @@ class CARESampler(dgl.dataloading.BlockSampler):
                 # extract each node from dict because of single node type
                 for node in seed_nodes:
                     edges = g.in_edges(node, form='eid', etype=etype)
-                    num_neigh = int(g.in_degrees(node, etype=etype) * self.p[block_id][etype])
+                    num_neigh = th.ceil(g.in_degrees(node, etype=etype) * self.p[block_id][etype]).int().item()
                     neigh_dist = self.dists[block_id][etype][edges]
+                    if neigh_dist.shape[0] > num_neigh:
                         neigh_index = np.argpartition(neigh_dist.cpu().detach(), num_neigh)[:num_neigh]
+                    else:
+                        neigh_index = np.arange(num_neigh)
                     edge_mask[edges[neigh_index]] = 1
                 new_edges_masks[etype] = edge_mask.bool()
 
@@ -56,6 +59,7 @@ class CAREConv(nn.Module):
         self.p = {}
         self.last_avg_dist = {}
         self.f = {}
+        # indicate whether the RL converges
         self.cvg = {}
         for etype in edges:
             self.p[etype] = 0.5
@@ -151,7 +155,7 @@ class CAREGNN(nn.Module):
     def RLModule(self, graph, epoch, idx, dists):
         for i, layer in enumerate(self.layers):
             for etype in self.edges:
-                if not layer.cvg:
+                if not layer.cvg[etype]:
                     # formula 5
                     eid = graph.in_edges(idx, form='eid', etype=etype)
                     avg_dist = th.mean(dists[i][etype][eid])
@@ -159,11 +163,17 @@ class CAREGNN(nn.Module):
                     # formula 6
                     if layer.last_avg_dist[etype] < avg_dist:
                         layer.p[etype] -= self.step_size
-                        layer.f.append(-1)
+                        layer.f[etype].append(-1)
+                        # avoid overflow, follow the author's implement
+                        if layer.p[etype] < 0:
+                            layer.p[etype] = 0.001
                     else:
                         layer.p[etype] += self.step_size
-                        layer.f.append(+1)
+                        layer.f[etype].append(+1)
+                        if layer.p[etype] > 1:
+                            layer.p[etype] = 0.999
+                    layer.last_avg_dist[etype] = avg_dist
 
                     # formula 7
-                    if epoch >= 10 and sum(layer.f[-10:]) <= 2:
-                        layer.cvg = True
+                    if epoch >= 9 and abs(sum(layer.f[etype][-10:])) <= 2:
+                        layer.cvg[etype] = True

python/dgl/data/fraud.py

@@ -109,7 +109,8 @@ class FraudDataset(DGLBuiltinDataset):
         
         data = io.loadmat(file_path)
         node_features = data['features'].todense()
-        node_labels = data['label']
+        # remove additional dimension of length 1 in raw .mat file
+        node_labels = data['label'].squeeze()
         
         graph_data = {}
         for relation in self.relations[self.name]:
@@ -118,8 +119,8 @@ class FraudDataset(DGLBuiltinDataset):
             graph_data[(self.node_name[self.name], relation, self.node_name[self.name])] = (row, col)
         g = heterograph(graph_data)
         
-        g.ndata['feature'] = F.tensor(node_features)
-        g.ndata['label'] = F.tensor(node_labels.T)
+        g.ndata['feature'] = F.tensor(node_features, dtype=F.data_type_dict['float32'])
+        g.ndata['label'] = F.tensor(node_labels, dtype=F.data_type_dict['int64'])
         self.graph = g
         
         self._random_split(g.ndata['feature'], self.seed, self.train_size, self.val_size)