Graph Sage OGBN Example with Perforated Backpropagation

[RorryB]

This PR is to include an example of how Perforated Backpropagation can be used to improve pytorch_geometric models. As well as adding a scheduler to the original example. Both of which improve upon the original results.

Run docker from torch_geometric directory

docker run --gpus all -i --shm-size=8g -v .:/pai -w /pai -t nvcr.io/nvidia/pyg:24.11-py3 /bin/bash

Within Docker

pip install -e .
cd examples
pip install PAI wheel file

Run original with:

CUDA_VISIBLE_DEVICES=0 python ogbn_train_original.py --dataset ogbn-products --batch_size 128

Results:

Test Accuracy: 75.52%

Run original scheduler with:

CUDA_VISIBLE_DEVICES=0 python ogbn_train_scheduler.py --dataset ogbn-products --batch_size 128

Results:

Test Accuracy: 77.51%

Run PAI with:

CUDA_VISIBLE_DEVICES=0 python ogbn_trainPAI.py --dataset ogbn-products --batch_size 128 --saveName ogbnPAI

Results:

Test Accuracy: 78.10%