Smaller gnn gives much stabler training

.gitignore

@@ -24,3 +24,4 @@ dataset/
 **.json
 **/checkpoint
 **/events.out.*
+**/test*

examples/lsc/pcqm4m/README.md

@@ -46,10 +46,10 @@ python main_mlpfp.py --log_dir $LOG_DIR --checkpoint_dir $CHECKPOINT_DIR --save_
 
 | Model              |Valid MAE  | Test MAE*   | \#Parameters    | Hardware |
 |:------------------ |:--------------   |:---------------| --------------:|----------|
-| GIN     | 0.1484 | 0.1621 | 11.6M  | GeForce RTX 2080 (11GB GPU) |
-| GIN-virtual     | 0.1382 | 0.1465 | 21.4M  | GeForce RTX 2080 (11GB GPU) |
-| GCN     | 0.1588 | 0.1710 | 5.9M  | GeForce RTX 2080 (11GB GPU) |
-| GCN-virtual     | 0.1459 | 0.1594 | 15.7M  | GeForce RTX 2080 (11GB GPU) |
+| GIN     | 0.1536 | 0.1678 | 3.8M  | GeForce RTX 2080 (11GB GPU) |
+| GIN-virtual     | 0.1396 | 0.1487 | 6.7M  | GeForce RTX 2080 (11GB GPU) |
+| GCN     | 0.1684 | 0.1838 | 2.0M  | GeForce RTX 2080 (11GB GPU) |
+| GCN-virtual     | 0.1510 | 0.1579 | 4.9M  | GeForce RTX 2080 (11GB GPU) |
 | MLP+Fingerprint     | 0.2044 | 0.2068 | 16.1M  | GeForce RTX 2080 (11GB GPU) |
 
 \* Test MAE is evaluated on the **hidden test set.**

examples/lsc/pcqm4m/main_gnn.py

@@ -12,6 +12,8 @@
 import argparse
 import time
 import numpy as np
+import random
+
 
 ### importing OGB-LSC
 from ogb.lsc import PygPCQM4MDataset, PCQM4MEvaluator
@@ -84,10 +86,10 @@ def main():
                         help='graph pooling strategy mean or sum (default: sum)')
     parser.add_argument('--drop_ratio', type=float, default=0,
                         help='dropout ratio (default: 0)')
-    parser.add_argument('--num_layers', type=int, default=7,
-                        help='number of GNN message passing layers (default: 7)')
-    parser.add_argument('--emb_dim', type=int, default=900,
-                        help='dimensionality of hidden units in GNNs (default: 900)')
+    parser.add_argument('--num_layers', type=int, default=5,
+                        help='number of GNN message passing layers (default: 5)')
+    parser.add_argument('--emb_dim', type=int, default=600,
+                        help='dimensionality of hidden units in GNNs (default: 600)')
     parser.add_argument('--train_subset', action='store_true')
     parser.add_argument('--batch_size', type=int, default=256,
                         help='input batch size for training (default: 256)')
@@ -106,6 +108,7 @@ def main():
     np.random.seed(42)
     torch.manual_seed(42)
     torch.cuda.manual_seed(42)
+    random.seed(42)
 
     device = torch.device("cuda:" + str(args.device)) if torch.cuda.is_available() else torch.device("cpu")
 

examples/lsc/pcqm4m/main_mlpfp.py

@@ -11,6 +11,7 @@
 import argparse
 import time
 import numpy as np
+import random
 
 from rdkit import Chem
 from rdkit.Chem import AllChem
@@ -147,6 +148,7 @@ def main_mlp():
     np.random.seed(42)
     torch.manual_seed(42)
     torch.cuda.manual_seed(42)
+    random.seed(42)
 
     device = torch.device("cuda:" + str(args.device)) if torch.cuda.is_available() else torch.device("cpu")