PyTextGCN
A re-implementation of TextGCN by Yao et al.: Graph Convolutional Networks for Text Classification. This implementation uses Cython for the text-to-graph transformation, making it rather fast. Graphs and GCN are based on the pytorch-geometric library.
Requirements
This project was built with:
- Python 3.8.5
- Cython 0.29.21
- CUDA 10.2 (optional for GPU support)
- scikit-learn 0.23.2
- pytorch 1.7.0
- torch-geometric 1.6.3
- gcc 9.3.0
- nltk 3.5
- scipy 1.5.2
At least the Text2Graph-module should work with other versions of these libraries, too.
Installation
From the project root, the cython compilation can be done with:
cd textgcn/lib/clib && python setup.py build_ext --inplace
Usage
To compute a graph from a list of strings (where each string contains the text of one document) called X, a list of labels called y and a list of test indices test_idx, simply run:
from textgcn import Text2GraphTransformer
t2g = Text2GraphTransformer()
graph = t2g.fit_transform(X, y, test_idx=test_idx)
The resulting object graph is a torch_geometric.data.Data object containing the resulting graph and can be processed by any torch-geometric-based network.
For more information on parameters of the Text2GraphTransformer and the resulting Data-object, consult the documentation in the source files.
Documentation
Currently resides in the source files.
How to reproduce our experiments
Run one of the Python-files in this folder. Each script will produce the result for one experiment for one seed. Seeds can be changed by editing the seed-variable in the first part of the file.