OT-Fusion of GCNs

Repository Structure

Here we describe the contents of the most important folders in this repository.

• src - Source code that performs the OT-Fusion. Parts of the code were inspired by the original Model Fusion via Optimal Transport.
• gnn_benchmarking - Slightly modified code of Benchmarking Graph Neural Networks used for defining, training and evaluating graph neural networks.
• results_reproduction - Scripts for reproducing the results from the project report.
• euler - Code and configs for setting up the parameter sweeps on Euler.
• models - Small pretrained models for experiments.

Reproducing Result

Reproducing Sweep Results

Results of sweeps are presented in sections 5.1 and 5.2 of the report. We recommend running the sweeps on Euler, as the number of configurations makes the process computationally heavy.

1. Setup Sweeps on Euler

Follow the instructions from gnn_fusion/euler/setup_instructions.md.

2. Evaluate Fused Models

After all the sweeps are finished, uncomment the first two lines in results_reproduction/evaluate_sweeps.sh, and execute (from gnn_fusiondirectory):

 sh results_reproduction/evaluate_sweeps.sh 

The above command should generate csv files with results in the gnn_fusion/results, print the contents of Table 1 in the terminal and save Figure 1 in gnn_fusion/report/figures folder.

Reproducing Remaining Tables

Results from sections 5.3, 5.4 and 5.5 can be easily reproduced on a laptop. If you want to setup the code on your laptop, follow all the steps in this section. If you have already configured Euler, you can continue in Euler terminal from step 3.

All commands should be executed from gnn_fusion root directory. We fixed the versions of the libraries in requirements.txt so that they work with Python 3.8. Adjust the versions accordingly for a newer Python.

1. Install Dependencies

Execute in the terminal:

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

2. Download ZINC Dataset

Execute in the terminal:

sh results_reproduction/download_molecules_dgl.sh

3. Reproduce Results

To reproduce results from section 5.3 execute:

sh results_reproduction/section_5_3.sh

To reproduce results from section 5.4 execute:

sh results_reproduction/section_5_4.sh

To reproduce results from section 5.5 execute:

sh results_reproduction/section_5_5.sh