2025-24679-HW1-Part2-image-autogluon-predictor

Duo or No Duo Image Classifier

Model Name: 2025-24679-HW1-Part2-image-autogluon-predictor
Author: Sebastian Andreu
Purpose: Binary image classification to detect whether “Duo” is present or absent in an image.
Intended Use: Educational purposes for practicing AutoML with neural networks and image classification.

Dataset

• Name: Duo or no Duo
• Source: Photos taken by Scotty McGee on an iPhone 15, converted from HEIC to RGB.
• Preprocessing: All images resized to 224 × 224 pixels.
• Splits:
  • Original: original images
  • Augmented: ≥300 synthetic images created with label-preserving transformations (random cropping, horizontal flips, rotations, color jitter, perspective distortion, slight blurring).
• Labels:
  • 1 → Duo present
  • 0 → Duo absent
• Features: Images only.
• Curator: Scotty McGee

Model

• Architecture: AutoML-selected image backbones via AutoGluon Multimodal (Timm models such as ResNet, MobileNet, EfficientNet).
• Training setup:
  • Training data: augmented split
  • Validation: subset of augmented split
  • Time budget: 600 seconds
  • Presets: medium_quality
  • Early stopping: enabled automatically by AutoGluon
• Hyperparameter search: AutoML over candidate image backbones, learning rates, batch sizes, and default AutoGluon optimization settings.

Metrics

• External validation: tested on original images.
• Accuracy: 1.0
• Weighted F1: 1.0

Limitations & Ethical Notes

• Dataset is small and synthetic augmentation may not fully represent real-world variations.
• Intended for educational purposes only, not for production or safety-critical use.
• All images are of non-sensitive objects, no PII.

AI Usage

• AutoGluon was used to automatically select model architecture, hyperparameters, and perform early stopping.
• Some code for dataset handling and visualization was assisted by GenAI.

License

• Educational / non-commercial use.