📊 CIFAR-10 Image Classification with Custom F-Beta Metric

1️⃣ Problem Statement and Goal of Project

The goal of this project is to classify images from the CIFAR-10 dataset into 10 categories while evaluating performance with a custom F-Beta score metric. This approach goes beyond simple accuracy, providing a more balanced evaluation that considers both precision and recall — essential in imbalanced or cost-sensitive classification problems.

2️⃣ Solution Approach

• Dataset Preparation:

  • Loaded CIFAR-10 dataset from tensorflow.keras.datasets.
  • Normalized pixel values to [0, 1].
  • One-hot encoded class labels.

• Custom Metric:

  • Implemented a FBetaScore class by subclassing tf.keras.metrics.Metric.
  • Uses built-in Precision and Recall metrics internally.
  • Prevents division-by-zero errors with epsilon handling.

• Model Architecture:

  • Convolutional layers:

    • Conv2D(32) → MaxPooling2D
    • Conv2D(64) → MaxPooling2D
    • Conv2D(128) → MaxPooling2D

  • Fully connected layers: Three Dense(50, ReLU) with Dropout(0.5).

  • Output layer: Dense(10, softmax).

• Training:

  • Optimizer: Adam
  • Loss: categorical crossentropy
  • Metrics: Accuracy + custom F-Beta score
  • Callback: EarlyStopping on validation loss with patience=3.

• Evaluation:

  • Computed accuracy and F-Beta score on test set.
  • Plotted F-Beta progression across epochs for both training and validation sets.

3️⃣ Technologies & Libraries

• Deep Learning Framework: TensorFlow / Keras
• Core Libraries: NumPy, Matplotlib
• Dataset Source: CIFAR-10 (tensorflow.keras.datasets)

4️⃣ Dataset Description

• Name: CIFAR-10
• Samples: 60,000 color images (32×32) in 10 classes, with 6,000 images per class.
• Split: 50,000 training, 10,000 test.
• Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck.

5️⃣ Installation & Execution Guide

# Clone the repository
git clone <your-repo-url>
cd <repo-name>

# Install dependencies
pip install tensorflow numpy matplotlib

# Open and run the notebook
jupyter notebook project02_fbeta.ipynb

6️⃣ Key Results / Performance

• Model achieves competitive accuracy and balanced F-Beta scores.
• F-Beta plots reveal convergence trends and potential class balance insights.
• EarlyStopping helps prevent overfitting while reducing training time.

7️⃣ Screenshots / Sample Output

(Refer to notebook for actual plots and logs)

• Model summary output from model.summary().
• Training/validation F-Beta score progression plot.

8️⃣ Additional Learnings / Reflections

• Implementing custom metrics in TensorFlow deepens understanding of the tf.keras.metrics API.
• F-Beta score is highly useful when precision and recall need to be balanced differently.
• Combining EarlyStopping with custom metrics allows for robust, efficient training.

---

👤 Author

Mehran Asgari 📧 imehranasgari@gmail.com 🐙 https://github.com/imehranasgari

---

📄 License

This project is licensed under the Apache 2.0 License – see the LICENSE file for details.

💡 Some interactive outputs (e.g., plots, widgets) may not display correctly on GitHub. If so, please view this notebook via nbviewer.org for full rendering.

---