Machine Learning — EfficientNetB0 Image Classifier (Transfer Learning)

A clean, reproducible implementation of an image classifier using EfficientNetB0 and transfer learning.
Works with any image dataset organized by class (not limited to flowers).

Results

Test set: 50 images (5 classes × 10).
Top-line metrics: Accuracy 1.0000, Macro-F1 1.0000 (from results/classification_report.txt).
Confusion matrix: perfectly diagonal (no misclassifications):

Example single-image prediction: model confidence per class; predicted label shown in the title.

Training curves (JSON): results/training_history_stage1.json, results/training_history_stage2.json, results/training_history_stage3.json.
During Stage 2–3, validation accuracy stays ≈1.00 with validation loss ≈0.27–0.33.

Note: The test set is intentionally small for demonstration. On larger or out-of-distribution datasets, results will vary; the evaluation script produces the same artifacts for reproducible reporting.

---

---

---

1) Overview

• Backbone: EfficientNetB0 (ImageNet-pretrained, loaded manually for robustness).
• Head: GlobalAveragePooling2D → Dropout(0.5) → Dense(num_classes, softmax).
• Preprocessing & Augmentation: In-model Keras augmentation (RandomFlip, RandomRotation(0.1), RandomZoom(0.1), RandomContrast(0.1)) and EfficientNet preprocess_input.
• Training schedule: 3 stages (head-only → partial unfreeze → full fine-tune) with ReduceLROnPlateau and EarlyStopping.
• Class imbalance: per-class sample weights derived from file counts in data/train.
• Artifacts: best checkpoint (.keras), final weights (.weights.h5), JSON training histories, evaluation report + confusion matrix, optional prediction visualizations.

---

2) Requirements

• Python 3.10+
• Install dependencies:

python -m venv venv
# Windows
venv\Scripts\activate
# macOS/Linux
source venv/bin/activate

pip install -r requirements.txt

---

3) Quick Start

Train (3 stages):

python -m src.train

Evaluate (classification report + confusion matrix):

python -m src.evaluate

Predict a single image (shows image + top-K bar chart; supports spaces in paths):

python -m src.predict "path/to/your image.jpg"
# save the visualization
python -m src.predict "path/to/your image.jpg" --save
# show more classes in the bar chart
python -m src.predict "path/to/your image.jpg" --topk 10 --save

Outputs & artifacts

• Best model: models/efficientnet_flower_model.keras
• Inference weights: models/efficientnet_weights.weights.h5
• Training histories: results/training_history_stage{1,2,3}.json
• Eval report: results/classification_report.txt
• Confusion matrix: results/confusion_matrix.png
• Prediction figures (optional): results/prediction_*.png

---

4) Model Configuration & Architecture

Input & preprocessing

• Input size: 224×224×3
• EfficientNet preprocess_input
• On-the-fly augmentation: RandomFlip, RandomRotation(0.1), RandomZoom(0.1), RandomContrast(0.1)

Backbone

• EfficientNetB0 without top, built with weights=None
• ImageNet no-top weights loaded manually via skip_mismatch=True
  (this avoids rare stem-conv shape mismatches on some TF/Keras versions while keeping ImageNet initialization)

Head

• GlobalAveragePooling2D → Dropout(0.5) → Dense(num_classes, activation="softmax")

Training — 3 stages

• Stage 1 (head-only): learning_rate = 1e-3, epochs = 8, backbone frozen
• Stage 2 (partial unfreeze): learning_rate = 1e-4, epochs = 15, unfreeze the last ~50 layers of the backbone
• Stage 3 (full fine-tune, optional): learning_rate = 5e-5, epochs = 5, unfreeze all backbone layers

Optimization & metrics

• Optimizer: Adam
• Loss: CategoricalCrossentropy(label_smoothing=0.05)
• Metrics: accuracy

Data pipeline

• batch_size = 32
• tf.data.AUTOTUNE prefetching

Class imbalance

• Class weights computed from file counts in data/train/*
• Injected as sample_weight per batch (works with one-hot labels)

Callbacks

• EarlyStopping(monitor="val_loss", patience=4, restore_best_weights=True)
• ReduceLROnPlateau(monitor="val_loss", factor=0.5, patience=2, min_lr=1e-6)
• ModelCheckpoint(..., save_best_only=True)

---

5) Using Other Datasets

This template is not restricted to flowers. It works with any multi-class image dataset as long as each class has its own folder under data/train, data/val, and (optionally) data/test. Class names are taken from folder names in data/train and must match across splits.

Recommended data split (effective practice)

• 70/20/10 (train/val/test) is a solid default for most small-to-medium projects.
  For larger datasets, 80/10/10 can be beneficial.
• Ensure the split is stratified per class so each class appears in all splits.
• Aim for at least ~50–100 images per class (more is better). If some classes are scarce, consider stronger augmentation and keep validation representative (do not oversample val).
• Keep the test set untouched until final evaluation.

---

6) Data Sources & Credits

• Example dataset used during development: “Tree Dataset of Urban Street Classification — Flower” on Kaggle (five well-represented flower classes were selected). The dataset is licensed under the GNU Lesser General Public License (LGPL), which permits use for academic and research purposes.

link: https://www.kaggle.com/datasets/erickendric/tree-dataset-of-urban-street-classification-flower

• EfficientNetB0: architecture background and official references by the original authors.