Skip to content

Latest commit

 

History

History
104 lines (77 loc) · 3.56 KB

README.md

File metadata and controls

104 lines (77 loc) · 3.56 KB
Raw

InfoGraph

中文版本 English Version

Background

Graph-level representations are essential in many real-world applications such as predicting the properties of molecules. InfoGraph is a recent state-of-the-art paper for learning graph-level representations through maximizing the mutual information between the graph-level representation and the representation of substructures of different scales.

Datasets

First, let us create a dataset root folder data under this application folder.

mkdir -p data && cd data

MUTAG

MUTAG is widely used dataset for graph learning. It consists of 188 organic molecules that were tested for mutagenicity. Assuming the root of your datasets is data, you can download this dataset using following command:

mkdir -p mutag/raw && cd mutag/raw
wget ftp://ftp.ics.uci.edu/pub/baldig/learning/mutag/mutag_188_data.can
wget ftp://ftp.ics.uci.edu/pub/baldig/learning/mutag/mutag_188_target.txt

PTC-MR

PTC-MR dataset has 344 molecules with a binary label indicating carcinogenicity of compounds on rodents. Assuming the root of your datasets is data, you can download this dataset using following command:

mkdir -p ptc_mr/raw && cd ptc_mr/raw
wget ftp://ftp.ics.uci.edu:21/pub/baldig/learning/ptc/ptc_MR_data.can
wget ftp://ftp.ics.uci.edu:21/pub/baldig/learning/ptc/ptc_MR_target.txt

After downloaed these datasets, run the script to preprocess the datasets:

python scripts/preprocess_data.py

Finally, the data folder looks like:

data
├── mutag
│   ├── processed
│   │   ├── data.npz
│   │   └── smiles.txt
│   └── raw
│       ├── mutag_188_data.can
│       └── mutag_188_target.txt
└── ptc_mr
    ├── processed
    │   ├── data.npz
    │   └── smiles.txt
    └── raw
        ├── ptc_MR_data.can
        └── ptc_MR_target.txt

Instructions

Configuration

The script unsupervised_pretrain.py is the entry for InfoGraph model to extract graph-level representation in an unsupervised manner. An example of hyperparameter configuration is model_configs/unsupervised_pretrain_config.json.

Training and Evaluation

For convenience, we provide a shell script scripts/unsupervised_pretrain.sh for easy experiments. Before running it, you need to manually configurate the number of runs, the root directory to datasets and the path to hyperparameter configuration.

runs=3
root="data"
config="model_configs/unsupervised_pretrain_config.json"

Evaluation results:

Datasets paper our run-1 our run-2 our run-3
MUTAG 89.01+/-1.13 91.43 90.20 90.45
PTC-MR 61.64+/-1.43 60.59 64.09 60.28

Reference

We refer to paper InfoGraph for the implementation, but the hyperparameters and network architecture might be slightly different.

InfoGraph

@inproceedings{sun2019infograph, title={InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization}, author={Sun, Fan-Yun and Hoffman, Jordan and Verma, Vikas and Tang, Jian}, booktitle={International Conference on Learning Representations}, year={2020} }