Merge branch 'PaddlePaddle:main' into pna

examples/pna/README.md

@@ -0,0 +1,37 @@
+# Principal Neighbourhood Aggregation for Graph Nets (PNA)
+
+[Principal Neighbourhood Aggregation for Graph Nets \(PNA\)](https://arxiv.org/abs/2004.05718)  is a graph learning model combining multiple aggregators with degree-scalers.
+
+
+### Datasets
+
+We perform graph classification experiment to reproduce paper results on [OGB](https://ogb.stanford.edu/). 
+
+### Dependencies
+
+- paddlepaddle >= 2.2.0
+- pgl >= 2.2.4
+
+### How to run
+
+
+```
+python main.py --config config.yaml   # train on ogbg-molhiv
+python main.py --config config_pcba.yaml # train on ogbg-molpcba
+```
+
+
+### Important Hyperparameters
+
+- aggregators: a list of aggregators name. ("mean", "sum", "max", "min", "var", "std")
+- scalers: a list of scalers name. ("identity", "amplification", "attenuation", "linear", "inverse_linear")
+- tower: The number of towers.
+- divide_input: hether the input features should be split between towers or not.
+- pre_layers: the number of MLP layers behind aggregators.
+- post_layers: MLP layers after aggregator.
+
+### Experiment results （ROC-AUC）
+|   | GIN   | PNA(paper result) | PNA(ours)|
+|-------------|----------|------------|-----------------|
+|HIV    | 0.7778  | 0.7905   | 0.7929     | 
+|PCBA   | 0.2266   | 0.2838   | 0.2801      |

examples/pna/config_pcba.yaml

@@ -40,6 +40,3 @@ log_filename: log.txt
 save_dir: ./checkpoints
 output_dir: ./outputs
 files2saved: ["*.yaml", "*.py", "./utils"]
-
-# python main_PCBA.py --type_net="complex" --batch_size=512 --lap_norm="none" --weight_decay=3e-6 --L=4 --hidden_dim=510 --out_dim=510 --residual=True 
-# --edge_feat=True  --readout=sum --graph_norm=True --batch_norm=True --aggregators="mean sum max" --scalers="identity" --config "configs/molecules_graph_classification_DGN_PCBA.json"  --lr_schedule_patience=4 --towers=5 --dropout=0.2 --init_lr=0.0005 --min_lr=0.00002 --edge_dim=16 --lr_reduce_factor=0.8