Graph-level representations using ensemble-based readout functions

The following is repository containing implementation of the paper "Graph-level representations using ensemble-based readout functions" by Jakub Binkowski, Albert Sawczyn, Denis Janiak, Piotr Bielak and Tomasz Kajdanowicz.

Citation

The paper is available on arXiv or in the proceedings of ICCS 2023.

  @InProceedings{10.1007/978-3-031-35995-8_28,
    author="Binkowski, Jakub and Sawczyn, Albert and Janiak, Denis and Bielak, Piotr and Kajdanowicz, Tomasz",
    editor="Miky{\v{s}}ka, Ji{\v{r}}{\'i} and de Mulatier, Cl{\'e}lia and Paszynski, Maciej and Krzhizhanovskaya, Valeria V. and Dongarra, Jack J. and Sloot, Peter M.A.",
    title="Graph-Level Representations Using Ensemble-Based Readout Functions",
    booktitle="Computational Science -- ICCS 2023",
    year="2023",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="393--405",
    isbn="978-3-031-35995-8"
    }

Experiments

Remarks

1. Experiment are stated in dvc.yaml file, and could be easly run with dvc repro command (but not in parallel, so it may take much time to finish)
2. Repeated experiments on several seeds are run with script: train_gnn_with_reps.py
3. Each experiment runs with corresponding configuration file experiments/confg/ensemble_readouts/hparams_<dataset>.yaml, which contains:
   • directories of the source dataset and further experiment output
   • hyperparameters, including model architecture specification

Installation

Repository relies on Python 3.10

• There are 4 files with requirements declared:

  • requirements-cpu.txt - packages only for CPU environments
  • requirements-gpu.txt - packages only for GPU environments
  • requirements.txt - packages common to GPU and CPU environments
  • requirements-dev.txt - linting tool to support code quality maintenance (optional)

• Install requirements for CPU environments:

   pip install -r requirements-cpu.txt -r requirements.txt -r requirements-dev.txt

• Install requirements for GPU environments:

  pip install -r requirements-gpu.txt -r requirements.txt -r requirements-dev.txt

Reproducing with DVC

To reproduce experiments with DVC, simply use the command below. Due to relatively low resource consumption, one might want to leverage parallel run described in the next section.

$ dvc pull data/datasets/{ENZYMES,MUTAG,REDDIT-MULTI-12K,ZINC}.dvc
$ dvc repro

Running parallel experiments

• Single experiment is relatively lightweight (consumes about ~10% of NVIDIA TITAN RTX)
• For the sake of fast computation of multiple experiments there is script, which helps to exploit resources
• To run experiments:
  • Ensure you have conda with environment called ensemble-readouts, containing all dependencies installed
  • Select dataset for which you want to run experiments, and look at the script run_training.sh
  • Run the scripts with parameters depending on available resources, e.g.,

    CUDA_VISIBLE_DEVICES=0 experiments/shell_runner/run_training.sh \
      --config-path experiments/config/deterministic_readouts/hparams_enzymes.yaml \
      --config-list experiments/config/config_names.yaml \
      --config-list-names ensemble_readouts \
      --num-jobs 4 \
      --accelerator gpu \
      --devices 1