xPerT: Extended Persistence Transformer

This repository contains code for training a GNN model (GIN) using PyTorch Geometric and generating extended persistence diagrams from graph datasets with giotto-deep.

Installation

Follow the steps below to set up the required environment:

git clone https://github.com/sehunfromdaegu/xpert.git
cd xpert
conda create --name xpert python=3.9
conda activate xpert
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
pip install -r requirements.txt
python -m pip install -U giotto-deep
conda install pyg=2.5.2 -c pyg

Optional: Fix scikit-learn errors

If you encounter any errors related to scikit-learn, you can reinstall the correct version as follows:

pip uninstall scikit-learn
pip install scikit-learn==1.1.1

Library modification

There are issues with specific versions of PyTorch Geometric and giotto-deep. To resolve these, apply the following fixes:

Locate the library path by running:

pip show torch_geometric

The library path will be listed under Location: PATH_TO_LIBS.

PyTorch Geometric

Open the file /PATH_TO_LIBS/torch_geometric/io/fs.py and modify line 193 as follows:

fs1.mv(path1, path2, recursive)  # Original
fs1.mv(path1, path2)             # Updated

giotto-deep

• Open the file /PATH_TO_LIBS/gdeep/data/datasets/persistence_diagrams_from_builder.py
• Modify line 206 to correct the graph labeling logic:

labels = (np.loadtxt(graph_labels, delimiter=",", dtype=np.int32).T + 1) // 2  # Original
labels = np.loadtxt(graph_labels, delimiter=",", dtype=np.int32).T             # Updated

Classification on Graph Datasets

To perform classification on graph datasets, use the following command:

python graph_classification.py --dataname <dataset> --model <modelname>

• dataset: Choose from the available options:

  • 'IMDB-BINARY'
  • 'IMDB-MULTI'
  • 'MUTAG'
  • 'PROTEINS'
  • 'COX2'
  • 'DHFR'

• modelname: Specify the model to be used:

  • 'xpert' (Extended Persistence Transformer)
  • 'gin' (Graph Isomorphism Network)
  • 'gin_assisted_concat' (GIN + xPerT by concat representations)
  • 'gin_assisted_sum' (GIN + xPerT by summing representations)

For example, to train the xPerT model on the MUTAG dataset, run:

python graph_classification.py --dataname MUTAG --model xpert

Classification on ORBIT dataset

To perform classification on ORBIT5K datasets, use the following command:

python orbit_classification.py

To perform classification on ORBIT5K datasets, use the following command:

python orbit_classification.py --samples_per_class 20000