multiview_gnn_refiner

Lightweight GNN module to refine 2-D keypoints by enforcing skeleton + cross-camera consistency.

Environment Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -c "import sys; sys.path.append('src'); print('OK')"

Dataset Folder Layout

Each dataset lives under ./data/{dataset_name}/. For example, the fly dataset is organized as:

data/
└── fly/
    ├── fly_ground_truth/
    ├── fly_ground_truth_OOD/
    ├── fly_predictions/
    └── fly_predictions_OOD/

These can simply be copy-pasted from the shared Drive

Running the GNN Refiner

python -m src.model.run_pipeline --config .\configs\pipeline.yaml

GPU Quickstart Guide

Item	Value
Project	uml-gnn-transfer
Instance name	gnn-gpu-1
Zone	us-east4-c
Image	pytorch-2-7-cu128-ubuntu-2404-nvidia-570-v20251013

1. Connect to the VM

From your Cloud Shell or local terminal (with gcloud SDK installed)

gcloud config set project uml-gnn-transfer
gcloud compute ssh gnn-gpu-1 --zone=us-east4-c

The first time you connect, GCP will generate an SSH key automatically. Type Y when prompted.

2. Verify GPU Access

Once inside the VM:

nvidia-smi
python3 -c "import torch; print(torch.cuda.is_available())"

You should see the NVIDIA T4 listed and True printed by PyTorch.

3. Set Up Your Environment

python3 -m venv ~/venv
source ~/venv/bin/activate
pip install --upgrade pip
git clone https://github.com/raavi02/UML-25.git
cd UML-25
pip install -r requirements.txt

4. Save Credits When Idle

Always stop the VM when you’re done:

gcloud compute instances stop gnn-gpu-1 --zone=us-east4-c

Restart later with:

gcloud compute instances start gnn-gpu-1 --zone=us-east4-c