Improvised Graph Neural Architecture Search Through Policy Optimization Methods

Try it yourself!

Setting up the codebase

Clone this repository on the local machine

git clone https://github.com/akshatowl/GraphNAS.git

After cloning, make sure all the dependencies are present. The suggested dependency versions are outdated and have depreciated functions. We installed every dependency around CUDA 12 instead, as per our GPU requirements.

If using CUDA 12, we suggest running this command to make sure the correct version of pytorch and torchvision are present compatible with CUDA 12:

pip3 install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

or, if you are using a conda environment:

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Requirements can be installed manually or by running the requirements.txt script:

pip install -r requirements.txt

or

pip3 install dgl scipy torch_scatter torch-cluster torch-sparse torch_geometric numpy hyperopt scikit_learn requests

Running the code

This is the main segment where reinforcement learning and our implementation will be reflected. The reinforcement learning algorithm can be specified using the --rlalgo tag, the default will be the baseline algorithm with the original discount function. This statement is run for designing an entire graph neural architecture. Running the baseline algorithm in the existing paper.

cd ~/GraphNAS
python -m graphnas.main --dataset Citeseer

Running the training with the Proximal Policy Optimization (PPO) Algorithm:

cd ~/GraphNAS
python -m graphnas.main --dataset Citeseer --rlalgo ppo

Running the training with the Trust Region Policy Optimization (TRPO) Algorithm:

cd ~/GraphNAS
python -m graphnas.main --dataset Citeseer --rlalgo trpo

After the model is constructed it will be saved in a folder directory within the project dependencies CiteSeer with a .pth extension.

Identify the model that is generated based on the time of generation and then copy the path of the model file.

In the semi directory in GraphNAS/semi/eval_designed_gnn.py enter the correct path of the model in the customer_architecture_path string.

custom_architecture_path = "/path_to_model/model.pth"

Then run the evaluation function to evaluate the training accuracy of the custom model with the Citeseer dataset.

cd ~/GraphNAS

Run the command:

python -m eval_scripts.semi.eval_designed_gnn

Experimentation test accuracy results

PPO and TRPO algorithms have discount generation using a 4 order butterworth filter with a cutoff frequency of 0.1 Hz. The results are generated for a Citeseer academic paper dataset.

Techniques	Test Accuracy
GraphNAS(Baseline)	73.7+/-0.2
APPNP	71.8+/-0.4
simple-NAS	71.7+/-0.6
GNAS with PPO	73.64+/-0.21
GNAS with TRPO	73.64+/-0.21