# PGNN

**Repository Path**: yh_cc/Class2019202001COMP620053

## Basic Information

- **Project Name**: PGNN
- **Description**: This is the project for P-GNN.
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2019-12-23
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

This repo contains the code for the Class 2019202001COMP620053.01.

It tries to reproduce the results reported by the paper [Position-aware Graph Neural Networks](http://proceedings.mlr.press/v97/you19b/you19b.pdf).
The author released their code in https://github.com/JiaxuanYou/P-GNN. We use their pre-processing code, 
since it is irrelevant to the idea itself. However, the container for 
the processed data is different. We rewrite the code by changing from the pytorch_geometric package to the python Natural Language Processing 
package fastNLP (https://github.com/fastnlp/fastNLP). 

Besides the reproduction, we also explore the effect of number of layers and 
the combination between P-GNN and GAT layers.

Therefore, our independent code work include:
* We adpat the pre-processing to use fastNLP dataset container. (About 150 lines of new code)
* We change the training process. (About 300 lines of new code)
* We adapt the model to work without Graph data container. (About 50 lines of new code)
* We tested this model in an extended setting. (About 150 lines of new code)
* We propose a new model named P-GAT. (About 200 line of new code.)
* We conduct over 1500 times experiments to get all the results.