Automatic Relation-aware Graph Network Proliferation

Contributions

We devise a RELATION-AWARE GRAPH SEARCH SPACE that comprises both node and relation learning operations. So, the ARGNP can leverage the edge attributes in some datasets, such as ZINC. It significantly improves the graph representative capability. Interestingly, we also observe the performance improvement even if there is no available edge attributes in some datasets.
We design a NETWORK PROLIFERATION SEARCH PARADIGM to progressively determine the GNN architectures by iteratively performing network division and differentiation. The network proliferation search paradigm decomposes the training of global supernet into sequential local supernets optimization, which alleviates the interference among child graph neural architectures. It reduces the spatial-time complexity from quadratic to linear and enables the search to thoroughly free from the cell-sharing trick.
The framework is suitable for solving node-level, edge-level, and graph-level tasks.

Relation-aware Graph Search Space

Network Proliferation Search Paradigm

Environment

Try the following command for installation.

# Install Graphviz
sudo apt install graphviz
# Install Python Environment
conda env create -f environment.yml
conda activate gnas

Download datasets

Some datasets (CLUSTER, TSP, ZINC, and CIFAR10) are provided by project benchmarking-gnns, the others (ModelNet10 and ModelNet40) are provided in Princeton ModelNet.

DATASET	TYPE	URL
CLUSTER	node	click here
TSP	edge	click here
ZINC	graph	click here
CIFAR10	graph	click here
ModelNet10	graph	click here
ModelNet40	graph	click here

Searching

We have provided scripts for easily searching graph neural networks on six datasets.

python start.py gpu=0 repeats=4 data=ZINC save='archs/start'

Training

We provided scripts for easily training graph neural networks searched by ARGNP.

CUDA_VISIBLE_DEVICES=0 python train.py ds=ZINC  optimizer=train_optimizer ds.load_genotypes='archs2/start/repeat3/ZINC/45/cell_geno.txt'

Visualization

We provided a python script for easily visualize searched architectures. Just run the following command.

python utils/plot_genotype.py

Cite

@InProceedings{Cai_2022_CVPR,
    author    = {Cai, Shaofei and Li, Liang and Han, Xinzhe and Luo, Jiebo and Zha, Zheng-Jun and Huang, Qingming},
    title     = {Automatic Relation-Aware Graph Network Proliferation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {10863-10873}
}