Merge branch 'rllm-team:main' into main

examples/rect.py

@@ -1,3 +1,18 @@
+# The RECT method from the
+# "Network Embedding with Completely-imbalanced Labels" paper.
+# ArXiv: https://arxiv.org/abs/2007.03545
+
+# RECT focuses on the zero-shot setting,
+# where only parts of classes have labeled samples.
+# As such, "unseen" classes are first removed from the training set.
+# Then we train a RECT (or more specifically its supervised part RECT-L) model.
+# Lastly, we use the Logistic Regression model to evaluate the performance
+# of the resulted embeddings based on the original balanced labels.
+
+# Datasets              Citeseer    |         Cora         | Pubmed
+# Unseen Classes  [1, 2, 5]  [3, 4] | [1, 2, 3]  [3, 4, 6] | [2]
+# RECT-L          66.50      68.40  | 74.80      72.20     | 75.30
+
 import argparse
 import copy
 import os.path as osp
@@ -11,27 +26,23 @@
 from rllm.datasets.planetoid import PlanetoidDataset
 from rllm.nn.models.rect import RECT_L
 
-# RECT focuses on the zero-shot setting,
-# where only parts of classes have labeled samples.
-# As such, "unseen" classes are first removed from the training set.
-# Then we train a RECT (or more specifically its supervised part RECT-L) model.
-# Lastly, we use the Logistic Regression model to evaluate the performance
-# of the resulted embeddings based on the original balanced labels.
 
 parser = argparse.ArgumentParser()
 parser.add_argument('--dataset', type=str, default='Cora',
                     choices=['Cora', 'CiteSeer', 'PubMed'])
 parser.add_argument('--unseen-classes', type=int, nargs='*', default=[1, 2, 3])
+parser.add_argument('--epochs', type=int, default=200, help="Training epochs")
 args = parser.parse_args()
 
 transform = T.Compose([
     T.NormalizeFeatures('l2'),
     T.SVDFeatureReduction(200),
-    T.GCNNorm()
+    T.GDC()
 ])
 
 path = osp.join(osp.dirname(osp.realpath(__file__)), '../data')
-dataset = PlanetoidDataset(path, args.dataset, transform=transform, force_reload=True)
+dataset = PlanetoidDataset(
+    path, args.dataset, transform=transform, force_reload=True)
 data = dataset[0]
 
 zs_data = T.RemoveTrainingClasses(args.unseen_classes)(copy.deepcopy(data))
@@ -48,18 +59,18 @@
 
 model, zs_data = model.to(device), zs_data.to(device)
 
-criterion = torch.nn.MSELoss(reduction='mean')
+criterion = torch.nn.MSELoss(reduction='sum')
 optimizer = torch.optim.Adam(model.parameters(), lr=0.001, weight_decay=5e-4)
 
 model.train()
 st = time.time()
-for epoch in range(1, 51):
+for epoch in range(1, args.epochs+1):
     optimizer.zero_grad()
     out = model(zs_data.x, zs_data.adj)
     loss = criterion(out[zs_data.train_mask], zs_data.y)
     loss.backward()
     optimizer.step()
-    # print(f'Epoch {epoch:03d}, Loss {loss:.4f}')
+    print(f'Epoch {epoch:03d}, Loss {loss:.4f}')
 et = time.time()
 model.eval()
 with torch.no_grad():

rllm/transforms/__init__.py

@@ -9,6 +9,8 @@
 from .knn_graph import KNNGraph  # noqa
 from .gcn_norm import GCNNorm  # noqa
 from .build_homo_graph import build_homo_graph  # noqa
+from .gdc import GDC
+
 
 general_transforms = [
     'BaseTransform',
@@ -23,6 +25,7 @@
     'RemoveSelfLoops',
     'KNNGraph',
     'GCNNorm',
+    'GDC',
 ]
 
 graph_builders = [