ResNet-18 transfer learning on the German Traffic Sign Recognition Benchmark (GTSRB) — 43 classes of real-world traffic signs. Built with PyTorch.

Traffic sign recognition is a core perception task in autonomous driving. A self-driving car must reliably classify signs under varying lighting, weather, angles, and occlusion — exactly the challenges present in the GTSRB dataset.

ResNet-18 (pretrained on ImageNet)
  └── Replace final FC layer:
      Linear(512) → Dropout(0.3) → Linear(256) → ReLU → Dropout(0.2) → Linear(43)

Full backbone fine-tuning (all layers trainable) with ImageNet-pretrained weights as initialization.

pip install torch torchvision matplotlib numpy

# Train with defaults (10 epochs)
python traffic_sign_classifier.py

# Custom settings
python traffic_sign_classifier.py --epochs 15 --batch_size 128 --lr 0.0005

Dataset downloads automatically (~600MB) on first run via torchvision.datasets.GTSRB.

• Transfer learning — ResNet-18 pretrained on ImageNet, fine-tuned on GTSRB
• Heavy augmentation — rotation, affine, color jitter, perspective (critical for real-world variance)
• Cosine annealing LR — smooth learning rate decay for better convergence
• ImageNet normalization — required for pretrained backbone compatibility
• Reproducibility — fixed seeds across all random generators
• Per-class analysis — identifies weak classes in an imbalanced dataset

The GTSRB dataset contains ~50,000 images of 43 German traffic sign classes. Key challenges:

• Images vary from 15×15 to 250×250 pixels
• Extreme lighting variation (day, night, shadows)
• Class imbalance (up to 10:1 ratio)
• Real-world noise, blur, and partial occlusion

├── traffic_sign_classifier.py     Training & evaluation pipeline
├── outputs/
│   ├── best_model.pth             Best model weights
│   ├── checkpoint.pth             Full training checkpoint
│   ├── 01_samples.png             Sample traffic signs
│   ├── 02_class_distribution.png  Class balance analysis
│   ├── 03_training_curves.png     Loss & accuracy curves
│   ├── 04_confusion_matrix.png    43×43 confusion matrix
│   ├── 05_wrong_predictions.png   Most confident errors
│   ├── 06_correct_predictions.png Correct prediction samples
│   └── 07_per_class_accuracy.png  Accuracy per sign type
└── README.md

• Python 3.10+
• PyTorch 2.x
• torchvision
• matplotlib, numpy

MIT