xprotocol/EfficientNet-B3-Cattle-Disease

EfficientNet-B3 — Cattle Disease Detection

Fine-tuned EfficientNet-B3 (ImageNet pre-trained) for multi-class cattle disease classification using a two-phase transfer learning strategy.

Classes

Index	Label
0	foot-and-mouth
1	healthy
2	lumpy

Model Architecture

• Backbone: EfficientNet-B3 (300×300×3 input)
• Head: GAP → BatchNorm → Dropout(0.3) → Dense(256, ReLU) → Dropout(0.2) → Softmax(3)
• Loss: Focal Loss (γ=2, α=0.25)
• Optimizer: AdamW + Cosine Annealing with Warm Restarts

Training Details

Parameter	Value
Input size	300×300
Phase 1 epochs	50 (frozen backbone)
Phase 2 epochs	30 (top-3 blocks unfrozen)
Phase 1 LR	1e-4
Phase 2 LR	1e-5
Weight decay	1e-4
Early stopping	patience=7 (val_loss)

Test Set Performance

Metric	Score
Accuracy	0.9482
Macro F1	0.9168
Macro AUC-ROC	0.9894

Usage (TensorFlow / Keras)

import keras

# Download the .keras file from the Hub and load:
model = keras.models.load_model(
    'efficientnet_b3_best.keras',
    custom_objects={
        'FocalLoss': FocalLoss,
        'EfficientNetPreprocess': EfficientNetPreprocess,
    }
)

# Predict (input: float32 numpy array of shape [N, 300, 300, 3] in [0, 255])
probs = model.predict(image_batch)   # shape (N, 3)
CLASS_NAMES = ['foot-and-mouth', 'healthy', 'lumpy']
predicted_class = CLASS_NAMES[probs.argmax(axis=1)[0]]

Dataset

Trained on devang03mgr/cattle-diseases-datasets.
Stratified split: 70% train | 15% val | 15% test.

Citation

If you use this model, please cite the original EfficientNet paper:

Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional
Neural Networks. ICML 2019. https://arxiv.org/abs/1905.11946