pydgn 0.7.0

A Python Package for Deep Graph Networks

PyDGN

Wiki

Description

This is a Python library to easily experiment with Deep Graph Networks (DGNs). It provides automatic management of data splitting, loading and the most common experimental settings. It also handles both model selection and risk assessment procedures, by trying many different configurations in parallel (CPU or GPU). This repository is built upon the Pytorch Geometric Library, which provides support for data management.

If you happen to use or modify this code, please remember to cite our tutorial paper:

Bacciu Davide, Errica Federico, Micheli Alessio, Podda Marco: A Gentle Introduction to Deep Learning for Graphs, Neural Networks, 2020. DOI: 10.1016/j.neunet.2020.06.006.

If you are interested in a rigorous evaluation of Deep Graph Networks, check this out:

Errica Federico, Podda Marco, Bacciu Davide, Micheli Alessio: A Fair Comparison of Graph Neural Networks for Graph Classification. Proceedings of the 8th International Conference on Learning Representations (ICLR 2020). Code

What's New

With PyDGN 0.7.0, we have released the first support for temporal experiments! In particular, we can now tackle node prediction in single graph sequence tasks! We rely on Pytorch Geometric Temporal for the definition of data and models.

Installation:

(We assume git and Miniconda/Anaconda are installed)

First, make sure gcc 5.2.0 is installed: conda install -c anaconda libgcc=5.2.0. Then, echo $LD_LIBRARY_PATH should always contain :/home/[your user name]/miniconda3/lib. Then run from your terminal the following command:

source setup/install.sh [<your_cuda_version>]
pip install pydgn

Where <your_cuda_version> is an optional argument that can be either cpu, cu102 or cu111 for Pytorch >= 1.8.0. If you do not provide a cuda version, the script will default to cpu. The script will create a virtual environment named pydgn, with all the required packages needed to run our code. Important: do NOT run this command using bash instead of source!

Remember that PyTorch MacOS Binaries dont support CUDA, install from source if CUDA is needed

Usage:

Preprocess your dataset (see also Wiki)

python build_dataset.py --config-file [your data config file]

Exampla

python build_dataset.py --config-file DATA_CONFIGS/config_PROTEINS.yml 

Launch an experiment in debug mode (see also Wiki)

python launch_experiment.py --config-file [your exp. config file] --splits-folder [the splits MAIN folder] --data-splits [the splits file] --data-root [root folder of your data] --dataset-name [name of the dataset] --dataset-class [class that handles the dataset] --max-cpus [max cpu parallelism] --max-gpus [max gpu parallelism] --gpus-per-task [how many gpus to allocate for each job] --final-training-runs [how many final runs when evaluating on test. Results are averaged] --result-folder [folder where to store results]

Example (GPU required)

python launch_experiment.py --config-file MODEL_CONFIGS/config_SupToyDGN_RandomSearch.yml --splits-folder DATA_SPLITS/CHEMICAL/ --data-splits DATA_SPLITS/CHEMICAL/PROTEINS/PROTEINS_outer10_inner1.splits --data-root DATA --dataset-name PROTEINS --dataset-class pydgn.data.dataset.TUDatasetInterface --max-cpus 1 --max-gpus 1 --final-training-runs 1 --result-folder RESULTS/DEBUG

To debug your code it is useful to add --debug to the command above. Notice, however, that the CLI will not work as expected here, as code will be executed sequentially. After debugging, if you need sequential execution, you can use --max-cpus 1 --max-gpus 1 --gpus-per-task [0/1] without the --debug option.

Grid Search 101

Have a look at one of the config files.

Random Search 101

Specify a num_samples in the config file with the number of random trials, replace grid with random, and specify a sampling method for each hyper-parameter. We provide different sampling methods:

• choice --> pick at random from a list of arguments
• uniform --> pick uniformly from min and max arguments
• normal --> sample from normal distribution with mean and std
• randint --> pick at random from min and max
• loguniform --> pick following the recprocal distribution from log_min, log_max, with a specified base

There is one config file, namely config_SupToyDGN_RandomSearch.yml, which you can check to see an example.

Data Splits

We provide the data splits taken from

Errica Federico, Podda Marco, Bacciu Davide, Micheli Alessio: A Fair Comparison of Graph Neural Networks for Graph Classification. Proceedings of the 8th International Conference on Learning Representations (ICLR 2020). Code

in the DATA_SPLITS folder.

Credits:

This is a joint project with Marco Podda (Github /Homepage), whom I thank for his relentless dedication.

Many thanks to Antonio Carta (Github /Homepage) for incorporating the Ray library (see v0.4.0) into PyDGN! This will be of tremendous help.

Many thanks to Danilo Numeroso (Github /Homepage) for implementing a very flexible random search! This is a very convenient alternative to grid search.

Many thanks to Alessio Gravina (Github /Homepage) for his invaluable help and expertise regarding the implementation of PyDGN temporal. We still have a lot of work to do!

Contributing

This research software is provided as-is. We are working on this library in our spare time.

If you find a bug, please open an issue to report it, and we will do our best to solve it. For generic/technical questions, please email us rather than opening an issue.

License:

PyDGN is GPL 3.0 licensed, as written in the LICENSE file.

Troubleshooting

As of 15th of August 2021, there is an issue with Pytorch 1.9.0 which impacts the CLI. This is why the setup script installs Pytorch 1.8.1 in the pydgn conda environment until Pytorch 1.10 is released (known to solve the issue).

--

If you get errors like /lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found:

• make sure gcc 5.2.0 is installed: conda install -c anaconda libgcc=5.2.0
• echo $LD_LIBRARY_PATH should contain :/home/[your user name]/[your anaconda or miniconda folder name]/lib
• after checking the above points, you can reinstall everything with pip using the --no-cache-dir option