-
Notifications
You must be signed in to change notification settings - Fork 0
Getting Started
The public repository contains the core Python package, adapters, configs, scripts, tests, Docker assets and technical notes. Raw datasets, large model checkpoints and generated experiment outputs are intentionally excluded.
Expected top-level components include:
adapters/
app/
configs/
docs/
incidentgraph/
scripts/
tests/
requirements.txt
requirements-deep.txt
pyproject.toml
Dockerfile
docker-compose.yml
git clone https://github.com/dranubhaparashar/IncidentGraph.git
cd IncidentGraph
python -m venv .venv
Set-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txtFor deep-learning dependencies:
pip install -r requirements-deep.txtThe repository also includes scripts/install_deep_windows.ps1; review it before execution when adapting the environment.
python -m pytest -qpython .\scripts\benchmark.py --help
python .\scripts\train_meva_activity.py --help
python .\scripts\train_meva_activity_resumable.py --help
python .\scripts\evaluate_meva_activity.py --helpThe codebase contains dedicated scripts for:
- downloading official annotations;
- downloading/preparing an annotated subset;
- inventorying annotations;
- building the activity manifest;
- rebalancing the split at video level when required;
- training and resumable training;
- evaluation and extended evaluation.
Read docs/REAL_MEVA_AND_DEEP_TRAINING.md before a long run.
The synthetic/system path uses configs/synthetic.yaml, scripts/generate_synthetic.py, scripts/benchmark.py, and the matching incidentgraph modules.
Warning
The controlled benchmark is a software/evaluation diagnostic, not a substitute for independent real multi-camera validation.
Keep raw video, private evidence, model checkpoints and generated outputs outside normal Git tracking. The repository .gitignore is intended to exclude these classes of files.
IncidentGraph · provenance preserved · uncertainty explicit · missing evidence stays unknown · consequential use requires human review