A browser-based user interface for the AVERINN neural network verification framework.
AVERINN is a formal verification tool for proving safety properties of artificial neural network controllers. The tool is powerful but currently exists only as a command line interface program. This project aims to create a more intuitive graphical user interface using a React-FastAPI stack, allowing researchers and engineers to configure, run, and visualize verifications on their own without having to interact directly with it through a terminal.
The project is being developed as part of a National Science Foundation Research Experiences for Undergraduates (REU) program at the University of New Mexico.
The browser interface aims to:
- Simplfy execution of AVERINN verification jobs
- Automatically create custom configuration files based on user selections
- Support multiple verification types/workflows
- Validate user inputs before execution
- Allow user to upload networks, dynamics, and spec sheets
- Display verification results in a clean, user-friendly format
- Reduce learning curve for new users of the framework
- React-based frontend
- FastAPI backend
- Automatic configuration file generation
- Support for dynamic and non-dynamic verification modes
- File uploads via frontend
- Parameter selection via frontend
- Safeguards that prevent unsupported files from being uploaded
- Execution of the AVERINN engine
- Result parsing and presentation
- React
- Javascript
- HTML
- CSS
- Bootstrap
- FastAPI
- Python
- AVERINN
- Sample networks, configs, and benchmark files
This project is currently under active development.
New features are being added as part of ongoing research into improving the usability of the verification tool.
These instructions assume you are starting from a fresh clone of the repository.
git clone ,repo-url>
cd <repo-name>
From the project root, create and activate a virtual environment:
python3 -m venv .venv
source .venv/bin/activate
Install the backend dependencies:
pip install --upgrade pip
pip install fastapi uvicorn python-multipart
pip install -r requirements.txt
Start the backend server:
cd backend
uvicorn main:app --reload
Open a second terminal and navigate to the frontend directory:
cd frontend
Install the Node dependencies:
npm install
Start the Vite development server:
npm run dev
The frontend should now be running locally. Vite will print the local address in the terminal. Copy that address into your browser to use the interface.
The backend and frontend need to run at the same time in separate terminals.
The node_modules/ folder is intentionally omitted from the repo. It will be recreated locally with
npm install
The Python virtual environment is also not included in the repo. It will be recreated locally with
python3 -m venv .venv
- Select verification type
- Note: only dynamic and non-dynamic verifications are active, more modes are being implemented
- Upload your desired network file and specifications sheet (and dynamics file, if applicable) before advancing
- Configure analysis options for the verification run
- Click "Start Verification" to run the tool with your selected settings and network file
Planned future improvements include:
- Rework "Selected Inputs" to reflect actual uploads
- Add model visualizer to upload card
Better visualization of results- Progress report during verification
Button lockout while tool is currently running- Options to run various tool types (linear, hybrid, etc)
- Improved error reporting
This interface is built on top of the existing AVERINN verification framework. I have not made any contributions to the engine itself.
This repo currently contains work developed for academic research.
Please consult the original AVERINN project (github.com/ratanlal02/AVERINN) regarding licensing before attempting to redistribute or incorporate the underlying framework into other projects.