Graph Neural Networks

Efficient PyTorch implementation for graph neural networks (GNN).

Dependencies

• PyTorch

• DGL==0.4.3 (Only used for automatically dataset downloading)

   pip3 install -r requirements.txt
  

Note that only DGL==0.4.3 will be produce correct results. Newer version will download a different dataset.

Job

Your job is to fill the class of APPNP to pass the test. To test locally, simply run:

  pytest

More specifically, change the forward definition of GraphAppnp in the file of "models/appnp.py".

class GraphAppnp(nn.Module):
    def __init__(self, alpha):
        super().__init__()
        self.alpha = alpha

    def forward(self, x, adj, h):
        '''
        You need to redefine the layer
        '''
        return x

You may also need to change the hyperparameters at Line53-Line57 in the file of "test_sample.py"

Training

GCN

• GCN: Graph Convolutional Network, ICLR 2017

• Training

  python3 train.py --model GCN --epochs 200 --lr 0.01 --dataset cora 
  python3 train.py --model GCN --epochs 200 --lr 0.01 --dataset pubmed
  python3 train.py --model GCN --epochs 200 --lr 0.01 --dataset citeseer
  

GAT

• GAT: Graph Attention Network, ICLR 2018

• Training

  python3 train.py --model GAT --epochs 100 --lr 0.01 --dataset cora 
  python3 train.py --model GAT --epochs 100 --lr 0.01 --dataset pubmed
  python3 train.py --model GAT --epochs 100 --lr 0.01 --dataset citeseer
  

SAGE

• GraphSAGE: Inductive Representation Learning on Large Graphs, NIPS 2017

• Training

  python3 train_sage.py --lr 0.01 --dataset cora
  python3 train_sage.py --lr 0.01 --dataset pubmed
  python3 train_sage.py --lr 0.01 --dataset citeseer
  

APPNP

• APPNP: Predict then Propagate: Graph Neural Networks Meet Personalized Pagerank, ICLR 2019

• Training

     python3 train.py --model APPNP --hidden 64 --epochs 100 --lr 0.01 --dataset cora
     python3 train.py --model APPNP --hidden 64 --epochs 100 --lr 0.01 --dataset pubmed
     python3 train.py --model APPNP --hidden 64 --epochs 100 --lr 0.01 --dataset citeseer
  

Usage

• train.py

  usage: train.py [-h] [--device DEVICE] [--model MODEL] [--data-root DATA_ROOT]
                  [--dataset DATASET] [--seed SEED] [--epochs EPOCHS] [--lr LR]
                  [--weight_decay WEIGHT_DECAY] [--hidden HIDDEN]
                  [--dropout DROPOUT]
  
  optional arguments:
    -h, --help            show this help message and exit
    --device DEVICE       cpu, cuda:0, cuda:1, etc.
    --model MODEL         GCN, GAT, or APPNP
    --data-root DATA_ROOT
                          dataset location
    --dataset DATASET     cora, citeseer, or pubmed
    --seed SEED           Random seed.
    --epochs EPOCHS       Number of epochs to train.
    --lr LR               Initial learning rate.
    --weight_decay WEIGHT_DECAY
                          Weight decay (L2 loss on parameters).
    --hidden HIDDEN       Number of hidden units.
    --dropout DROPOUT     Dropout rate.
  

• train_sage.py

  usage: train_sage.py [-h] [--device DEVICE] [--data-root DATA_ROOT]
                       [--dataset DATASET] [--save SAVE] [--aggr AGGR] [--lr LR]
                       [--factor FACTOR] [--min-lr MIN_LR] [--patience PATIENCE]
                       [--batch-size BATCH_SIZE] [--epochs EPOCHS]
                       [--early-stop EARLY_STOP] [--weight_decay WEIGHT_DECAY]
                       [--seed SEED]
  
  optional arguments:
    -h, --help            show this help message and exit
    --device DEVICE       cpu, cuda:0, cuda:1, etc.
    --data-root DATA_ROOT
                          learning rate
    --dataset DATASET     cora, citeseer, pubmed
    --save SAVE           model file to save
    --aggr AGGR           Aggregator: mean, pool, or gcn
    --lr LR               learning rate
    --factor FACTOR       EarlyStopScheduler factor
    --min-lr MIN_LR       minimum lr for EarlyStopScheduler
    --patience PATIENCE   patience for Early Stop
    --batch-size BATCH_SIZE
                          number of minibatch size
    --epochs EPOCHS       number of training epochs
    --early-stop EARLY_STOP
                          number of epochs for early stop training
    --weight_decay WEIGHT_DECAY
                          Weight decay (L2 loss on parameters).
    --seed SEED           Random seed.