WenDEVLIFE/KoalasModel

Release v1.0.0

Date: September 19, 2026

What's New

Initial release of Dog Emotion Detector — a mobile-first classifier for detecting dog emotions from images.

Features

• 5-class emotion classification: alert, angry, frown, happy, relax
• Transfer learning: EfficientNetV2B0 backbone with ImageNet weights
• TFLite export: 24 MB model with ~22ms CPU inference
• Full pipeline: data loading → augmentation → training → evaluation → export
• Configurable: YAML-based configuration for model, training, and export

Performance

Metric	Value
Accuracy	69.9%
Macro F1	69.2%
Model size	24 MB
Inference latency	~22ms (CPU)

Per-class Results

Class	F1 Score
happy	0.89
relax	0.79
frown	0.61
angry	0.53
alert	0.58

Installation

pip install -e .

Usage

python main.py train --config configs/config.yaml

Model Architecture

• Backbone: EfficientNetV2B0 (ImageNet pretrained)
• Head: GlobalAveragePooling2D → Dense(256, relu) → Dropout(0.5) → Dense(5, softmax)
• Optimizer: Adam (lr=0.0003)
• Loss: Categorical crossentropy with label smoothing (0.1)
• Regularization: Early stopping (patience=20), reduce-LR-on-plateau (patience=8, factor=0.5)

Dataset

• Total images: 9,325
• Classes: 5 (alert, angry, frown, happy, relax)
• Images per class: ~1,865
• Split: 80% train / 10% validation / 10% test

Known Limitations

• Angry/alert/frown classes have lower F1 due to subjective visual similarity
• Performance is limited by dataset size (~1.8K images per class)
• Further improvement requires more data or advanced augmentation (CutMix/Mixup)

Future Improvements

• CutMix/Mixup augmentation for better generalization
• Test-time augmentation for inference boost
• Ensemble methods for production deployment
• Larger training dataset (50K+ images)

License

MIT

---

license: apache-2.0