Deep Safety

This is a repository to study safety concepts for deep learning in safety-critical applications. The repository is part of the lecture at DHBW Ravensburg.

This course includes a project whose goal is to validate a machine learning based system for traffic sign classification that is assumed to be part of a self-driving vehicle. During the course, you will you work on several assignments that cover specific aspects of safety for deep learning and prepare you for the final project.

Quickstart

1. Clone the repository:

   git clone [email protected]:schutera/DeepSafety.git

2. Create a new virtual environment and install dependencies.

   Either using Conda (recommended):

   conda env create --file environment.yaml

   or using venv:

   python -m venv deepsafety_env
   source deepsafety_env/bin/activate
   pip install -r requirements.txt

   Please refer to the PyTorch Get Started guide for installation details if you want to use a GPU.

3. Activate your virtual environment, e.g., Conda:

   conda activate deepsafety_env

4. Train a new model using the default parameters:

   python -m train

   If you want to use custom parameters, you can use the -h argument to get an overview over the possible arguments.

5. Launch the MLFlow UI to view your logged run in the tracking UI. From a terminal in the repository root directory run:

   mlflow ui --port 8080 --backend-store-uri sqlite:///mlruns.db

   Then, navigate to http://localhost:8080 in your browser to view the results.

6. Evaluate your trained model:

   python -m eval --data-dir <path_to_evaluation_data> --run-id <MLFlow_run_id>