Advanced Deep Learning Techniques: Custom Training Loops and Model Optimization

This project demonstrates a practical and in-depth understanding of low-level TensorFlow and Keras APIs by implementing custom training loops for two distinct image classification tasks. It begins with a foundational CNN training pipeline on the Eurosat dataset using tf.GradientTape and progresses to an optimized MLP for MNIST classification with advanced performance-enhancing techniques.

---

1. Problem Statement and Goal of Project

High-level APIs such as model.fit() abstract much of the training process, which is convenient but limits flexibility when implementing custom logic, advanced optimizers, or non-standard evaluation metrics. This project addresses that limitation by manually managing the training process, offering complete control over each step.

Key objectives:

1. Foundational Implementation: Train a CNN from scratch on Eurosat satellite images with a custom GradientTape loop.
2. Performance Optimization: Apply architectural and training improvements to a MNIST classifier to enhance accuracy and robustness.

---

2. Solution Approach

Two Jupyter Notebooks are provided to illustrate the transition from basic custom training to optimized, production-ready training.

Notebook 1 – Eurosat Classification

File: simple_just_for_learning_not_metric.ipynb

• Dataset Handling: Loaded Eurosat from tensorflow_datasets; split into 70% train, 15% validation, 15% test.

• Data Pipeline: Built with tf.data for efficient batching, shuffling, and prefetching.

• Preprocessing:

  • Resize to 64×64
  • Normalize to [0, 1]
  • One-hot encode labels

• Augmentation: Random flips, rotations, zooms, contrast adjustments (applied on-the-fly).

• Model: CNN architecture implemented manually.

• Training Loop: Implemented with tf.GradientTape, including:

  • Gradient computation & manual weight updates
  • Validation monitoring
  • Model checkpoint saving
  • Early stopping

• Monitoring: Integrated TensorBoard for loss/accuracy visualization.

---

Notebook 2 – MNIST Classification with Optimization

File: low_level_api_better_acc.ipynb

• Dataset Handling: Loaded MNIST via keras.datasets.

• Loss Function: SparseCategoricalCrossentropy (works with integer labels, memory-efficient).

• Architecture: Multi-Layer Perceptron with:

  • Batch Normalization (stabilizes and accelerates training)
  • Dropout (reduces overfitting)

• Training Loop Enhancements:

  • Early Stopping (manual implementation)
  • ReduceLROnPlateau (learning rate adjustment on validation loss plateau)

• Evaluation:

  • Test set accuracy
  • Confusion Matrix for per-class performance

---

3. Technologies & Libraries

• Frameworks: TensorFlow, Keras
• Libraries: TensorFlow Datasets, NumPy, Matplotlib, scikit-learn, Tqdm
• Tools: Jupyter Notebook, TensorBoard

---

4. Description about Dataset

• Eurosat: 27,000 labeled satellite images (64×64 px, RGB) in 10 land use classes (e.g., Forest, River, Industrial).
• MNIST: 70,000 grayscale handwritten digits (28×28 px, classes 0–9).

---

5. Installation & Execution Guide

# 1. Clone the repository
git clone https://github.com/imehranasgari/your-repo-name.git
cd your-repo-name

# 2. Install dependencies
pip install -r requirements.txt

# 3. Launch Jupyter Notebook
jupyter notebook

Open either:

• simple_just_for_learning_not_metric.ipynb
• low_level_api_better_acc.ipynb

---

6. Key Results / Performance

• Eurosat Notebook:

  • Robust, reusable training pipeline with augmentation and monitoring.
  • Demonstrates complete manual training loop.

• MNIST Notebook:

  • Significant accuracy improvement with Batch Normalization, Dropout, and LR scheduling.
  • Clear per-class breakdown via confusion matrix.

---

7. Screenshots / Sample Output

(Use your own prepared screenshots for clarity — examples include:)

• Eurosat dataset sample images
• MNIST training/validation curves
• MNIST confusion matrix

---

8. Additional Learnings / Reflections

• Moving beyond model.fit() provided deeper insight into:

  • Gradient descent & backpropagation
  • Metric calculation
  • Manual control over optimization flow

• The first notebook emphasized pipeline building; the second showcased model optimization for higher accuracy.

• Some notebooks intentionally use simpler models or achieve lower metrics — these are learning exercises, not production constraints.

---

👤 Author

Mehran Asgari Email: imehranasgari@gmail.com GitHub: 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.

---