This repository contains code to reproduce the results on the synthetic and public datasets.
Here we share an installation guide for Mac OS X and Ubuntu. Note that, due to a bug, AutoML will not work in Mac OS X (automl/auto-sklearn#1591).
First create a python environment
conda create -n gsim python=3.7
and activate it
conda activate gsim
then install all the required packages
pip install -r requirements.txt
make sure to install tensorflow separately, we use this to generate the tensorboards which display the results
pip install tensorflow
and finally, we need to install the POT library to calculate the Wasserstein distance. Optionally, to speed things up, you can install pathos for multi-core processing.
pip install --user -U https://github.com/PythonOT/POT/archive/master.zip
pip install pathos
pip install p_tqdm
Please follow the installation steps for Ubuntu, but install Tensorflow as follows.
Make sure to install tensorflow, for M1 macs run:
python -m pip install tensorflow-macos
pip install tensorflow-metal
Then install all other requirements:
Run pip install -r requirements.txt
The current version of POT, 0.7, does not yet include some important bug fixes. Using the latest version is therefore recommended. This is done via
pip install --user -U https://github.com/PythonOT/POT/archive/master.zip
If you want to calculate the Wasserstein distances in parallel, you need the Pathos package: pip install pathos, or for a pretty progress bar p_tqdm: pip install p_tqdm which uses Pathos in the background.
For Mac OS X we have the following bug: automl/auto-sklearn#1591
We included various .sh files, these list the experiments you could do.
- experiments_public.sh: run the experiments on the public dataset with svm classifier
- experiments_public_automl.sh: see above, but we use automl
- experiments_public_random.sh: we use a random set of reference graphs
- experiments_synthetic_different_classes.sh: classifies different classes
- experiments_synthetic_same_class.sh: classifies graphs from the same class, but with different settings
You can define your own experiments, execute the following command for guidance
python experiments.py -h
to get the help menu, or modify the below example command
python experiments.py -d https://ls11-www.cs.tu-dortmund.de/people/morris/graphkerneldatasets/MUTAG.zip --online -i 8 -r 128 --categorical --model wwlr --n_jobs 12 --gridsearch
To validate our results, we use publicly available dataset and synthetic data. Here we describe how to generate the synthetic data.
The following command stores the generated datasets in the data/ folder
python generate_fake_data.py