shuowang-ai/geniepath-pytorch

This is a PyTorch implementation of the GeniePath model in <GeniePath: Graph Neural Networks with Adaptive Receptive Paths> (https://arxiv.org/abs/1802.00910)

GeniePath-pytorch

This is a PyTorch implementation of the GeniePath model in GeniePath: Graph Neural Networks with Adaptive Receptive Paths

GeniePath, a scalable approach for learning adap- tive receptive fields of neural networks defined on permuta- tion invariant graph data. In GeniePath, we propose an adap- tive path layer consists of two complementary functions de- signed for breadth and depth exploration respectively, where the former learns the importance of different sized neighbor- hoods, while the latter extracts and filters signals aggregated from neighbors of different hops away. Our method works in both transductive and inductive settings, and extensive ex- periments compared with competitive methods show that our approaches yield state-of-the-art results on large graphs

Model

Requirements

• pyG

Usage

Install packages

pip install -r requirements.txt
# pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

Choose Model

in ppi_geniepath.py

# from model_geniepath import GeniePath as Net
from model_geniepath import GeniePathLazy as Net

Run

python ppi_geniepath.py

TODO

• Finish the rough implementation, f1_score: 0.9709 for GeniePath, 0.9762 for GeniePathLazy (dim = 256, lstm_hidden = 256).
• Tune the model

Reference

• Contribute to pytorch_geometric/examples/geniepath.py

geniepath-pytorch's Issues

loss NAN after the first epoch

After the first epoch both your model result in NAN loss. I am using P100 and havs anyone encountered this problem before?

为什么使用“from model_geniepath import GeniePath as Net”作为网络模型，只能得到0.45的f1-score?

实际运行时，
使用“from model_geniepath import GeniePath as Net”作为网络模型，只能得到0.45的f1-score。
而README.md说能达到0.9709，请问是需要设置一些参数吗？谢谢

为什么GAT的PPi的结果和文章里差那么多

这里的PPI设定和GAT文章里的一样吗？为什么GAT里PPI能达到97.3，这篇文章只报了81？

where is the skip connection in the "GeniePath-lazy" model ?

hello, thanks for your great work !
I am wondering where is the skip connection in the "GeniePath-lazy" model ?

where is the "Efficient Numerical Computation" ?

I can't find any code of the part about "Efficient Numerical Computation" proposed in your paper.