Contributing to ProG-V2

Thanks for contributing to ProG-V2. This guide describes the public development workflow, coding conventions, and validation steps expected for pull requests.

Development Setup

Use Python 3.9 or 3.11. Python 3.11 is recommended for local development.

conda create -n prog-v2 python=3.11 -y
conda activate prog-v2
pip install -e ".[dev]"
pre-commit install

Quick environment check:

python -c "import prompt_graph; print(prompt_graph.__file__)"
pytest tests/test_factory.py -v

Branches and Commits

Use one branch per goal.

Prefix	Use case	Example
fix/	Bug fixes	fix/relief-node-eval-device
feat/	New user-facing functionality	feat/new-prompt-strategy
refactor/	Internal structure changes	refactor/strategy-optimizer-setup
test/	Test additions or fixes	test/backbone-registry
docs/	Documentation updates	docs/public-results
chore/	Maintenance	chore/update-ci

Commit messages should use the matching prefix:

fix: handle WebKB multi-split train masks
docs: publish GCN benchmark report
test: cover strategy registration

Explain why a change is needed in the commit body when the reason is not obvious.

Coding Conventions

Paths

Do not hardcode ./data or ./Experiment in runtime code. Use helpers from prompt_graph.utils.paths, such as:

• excel_result_dir
• sample_dir
• pretrained_model_dir
• induced_graph_dir

Device handling

Use prompt_graph.utils.resolve_device(device) instead of constructing torch.device(...) directly in task setup code.

Logging

Use the project logger for intermediate state:

from prompt_graph.utils import get_logger

logger = get_logger(__name__)
logger.info("epoch %d loss=%.4f", epoch, loss)

Keep print() for user-facing final results only.

Type and compatibility notes

The project supports Python 3.9 and 3.11. Avoid runtime-only syntax that breaks Python 3.9.

Running Checks

Before opening a pull request, run:

ruff check .
ruff format --check .
pytest tests/ -v

For faster iteration:

pytest tests/test_factory.py -v
pytest tests/test_strategy_registry.py -v
pytest tests/test_strategy_gpf.py -v

Adding a Prompt Strategy

1. Add a strategy implementation under prompt_graph/tasker/strategies/.
2. Register it with @register_strategy("PromptName").
3. Import the module in prompt_graph/tasker/strategies/__init__.py so the registry is populated on package import.
4. Add or update initialization logic if the prompt needs custom modules.
5. Add a smoke test under tests/.

At minimum, a new strategy should be able to run a small Cora or MUTAG smoke configuration.

Adding a Backbone

1. Add the model implementation or wrapper under prompt_graph/model/.
2. Register it in the model factory used by build_gnn.
3. Add a construction + forward smoke test in tests/test_factory.py.
4. Verify at least one pretrain/downstream path with the new backbone.

Adding a Dataset

1. Add loader support in prompt_graph/data/load4data.py.
2. Update dataset lists in prompt_graph/defines.py.
3. Add or update a data-loader smoke test.
4. Ensure generated files use prompt_graph.utils.paths helpers.

Benchmark Results

Public merged reports should live under results/. Keep raw local outputs, temporary merge workspaces, and machine-specific logs out of the repository.

The current public benchmark reports are:

results/benchmark-gcn/          # node- and graph-classification report
results/link-prediction-gcn/    # edge task (link prediction) report

Each contains a flat summary.csv, a final_matrices.xlsx workbook, and bench-style per-dataset Excel matrices.

Pull Request Checklist

• The PR has one clear goal.
• New runtime code avoids hardcoded paths.
• New task/model code uses centralized device handling.
• New intermediate output uses logging instead of print().
• Tests were added or updated for new behavior.
• ruff check . passes.
• ruff format --check . passes.
• Relevant pytest targets pass.
• Public docs/results do not include local machine or private execution metadata.