GraphPoolingGarden

A repo for baseline of graph pooling methods.

中文

Datasets

• TuDataset
  • D&D
  • PROTEINS
  • ENZYMES
  • NCI1/NCI109
  • Reddit-Binary
• OGB
  • ogbg-molhiv
  • ogbg-ppa
  • ogbg-code2

Pooling methods

• set2set
• sagpool(sequence/hierarchical)
• graph-U-net
• Diffpool

Model Framework

• hierarchical model
• sequencial model
• U-net-like model: only for graph U-net model
• diffpooling model: only for diffpool model

Readout methods

• mean
• max
• sum
• set2set

ConvLayer

• GCN: layers/gcn_layer.py
• GIN: layers/gin_layer.py
• GraphSAGE:layers/graphsage_layer.py

Usage

1. create the config json file

create a config.json in configs folder, an example is like this:

{
    "dataset_name": ["ENZYMES"],
    "batch_size": 2,
    "epochs": 100,
    "seed": [1,2,3,4,5],
    "model":"global",
    "gnn_type": "gcn",
    "num_layer": 3,
    "emb_dim": 300,
    "drop_ratio": 0.5,
    "virtual_node": "False",
    "residual": "False",
    "JK": "last",
    "pooling": "sagpool",
    "sagpool": {
        "keep_ratio": 0.8,
        "activation": "tanh",
        "layer_num": 1
    }
}

When the key is a list such as dataset_name, all permutation will be trained and log in the csv file named as the comabination of the key values of 'dataset_name' and 'pooling'. In the examle, the final result will be saved in ENZYMES_sagpool.csv.

Warn:

Not all values can be written list-like, only the following keys can:

• dataset_name
• seed

2. running train.py to train, valid and test

python graphpoolinggarden/train.py --config configs/config.json

Acknowledgement

The program is built based on the code of the following released codes and programs:

• Open Graph Benchmark(OGB)
• Graph-U-Nets
• DiffPool
• DGL
• Pytorch Geometrics

We appreciate the authors' effort for the contribution to the research community.