PyTorch Pets Classifier

Minimal, production-style baseline for multi-class image classification on the Oxford-IIIT Pets dataset (37 breeds). Uses transfer learning with torchvision ResNet18 ImageNet weights and runs on macOS MPS or CPU.

Setup

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

Train

python src/train.py --config configs/default.yaml

Common overrides:

python src/train.py --epochs 10 --batch-size 64 --lr 3e-4 --freeze-epochs 2 --num-workers 0

Best checkpoint is saved to ./checkpoints/best.pt.

Evaluate

python src/eval.py --ckpt checkpoints/best.pt

Predict

python src/predict.py --ckpt checkpoints/best.pt --image path/to/image.jpg

Example output:

Top-1: abyssinian (0.9234)
Top-5:
	abyssinian (0.9234)
	bengal (0.0345)
	siamese (0.0121)
	ragdoll (0.0098)
	birman (0.0076)

macOS MPS

The code automatically selects MPS if available via torch.backends.mps.is_available(). If MPS is not available, it falls back to CPU.

Repo hygiene

• Dataset downloads to ./data (not committed)
• Checkpoints saved to ./checkpoints (not committed)

Roadmap

• Grad-CAM visualization
• AMP training
• Optuna hyperparameter search
• Weights & Biases logging
• ONNX export