Download Dataset

Install dependencies:

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

Create data/.env:

ROBOFLOW_API_KEY=your_key

Download the dataset:

python data/download.py

Expected dataset path after download:

data/Self-Driving-Car-3/export/images
data/Self-Driving-Car-3/export/labels

Run Baseline Training

Train the baseline from configs/baseline.yaml:

python -m src.train --config configs/baseline.yaml

Train with a custom run name:

python -m src.train --config configs/baseline.yaml --name baseline

Training creates a run directory in runs/:

runs/YYYYMMDD_HHMMSS/
runs/YYYYMMDD_HHMMSS_name/

After training, the run directory should contain baseline artifacts such as:

config.yaml
result.json
weights/best.pt
weights/last.pt

An example saved training run is available in runs.example/yolov8n_baseline with args.yaml, results.csv, and weights/.

Run Evaluation

src.train already runs validation and test evaluation after training and stores the metrics in result.json.

You can also run evaluation separately for a saved checkpoint:

python -m src.eval \
  --config configs/baseline.yaml \
  --model runs/<run_name>/weights/best.pt

Run evaluation on only part of the test split and save ground truth plus predictions:

python -m src.eval \
  --config configs/baseline.yaml \
  --model runs/<run_name>/weights/best.pt \
  --split test \
  --limit 64 \
  --save-gt-pred

This writes result.json, <split>_subset.txt, gt.json, and pred.json into the run directory.

Run Inference

With config:

python -m src.inference --config configs/baseline.yaml

Direct run with local model:

python -m src.inference \
  --model models/best_yolo_auto_trasnport.pt \
  --source data/Self-Driving-Car-3/export/images

Results are written to runs/.