Skip to content

Latest commit

 

History

History
143 lines (89 loc) · 4.22 KB

README.md

File metadata and controls

143 lines (89 loc) · 4.22 KB

DL_Keras_Model_Subclassing — Custom Models with TensorFlow/Keras

1) Project Title

Deep Learning with Keras Model Subclassing — Custom Architectures, Loss Functions, and Training Loops

2) Problem Statement and Goal of Project

This project demonstrates full customization of deep learning models in TensorFlow/Keras using the Model Subclassing API. The aim is to:

  • Build custom model architectures by directly subclassing tf.keras.Model.
  • Implement custom layers, user-defined loss functions, and forward passes with complete flexibility.
  • Showcase advanced training workflows and model control beyond the standard Sequential/Functional APIs.

3) Solution Approach

The notebook implements:

  1. Custom Model Definition

    • Inherits from tf.keras.Model.
    • Defines layers inside __init__ and the forward computation in call(...).

  2. Custom Loss Function

    • Implemented directly in Python with TensorFlow operations.

  3. Flexible Training

    • Uses the Keras .compile() / .fit() workflow but allows custom training logic if needed.

  4. MNIST Classification Example

    • Uses TFDS to load, normalize, and one-hot encode the dataset.
    • Trains the custom model on 28×28 grayscale images.

  5. Model Diagram

    • Created using keras.utils.plot_model(...) to visualize the architecture.

This approach gives full control over architecture design and training behavior, enabling experimentation that’s difficult with high-level APIs.

4) Technologies & Libraries

  • Python
  • TensorFlow / Keras
  • TensorFlow Datasets (TFDS)
  • NumPy
  • pydot / graphviz (for architecture visualization)

5) Description about Dataset

  • MNIST (from TFDS)

    • Handwritten digits (0–9).
    • Input shape: (28, 28, 1) grayscale images.
    • Labels: One-hot encoded with depth = 10.

  • Data processing steps in code:

    • Normalization (tf.divide(..., 255.0))
    • One-hot encoding (tf.one_hot(..., depth=10))
    • Shuffling, batching, and prefetching with tf.data.

6) Installation & Execution Guide

Prerequisites

  • Python 3.x
  • TensorFlow (GPU optional, but recommended)

Install (pip)

pip install tensorflow tensorflow-datasets numpy pydot graphviz

On some systems you must also install Graphviz binaries: sudo apt-get install graphviz

Run

  1. Open model_subclassing_me.ipynb in Jupyter/VS Code.

  2. Execute cells sequentially.

  3. The notebook will:

    • Download MNIST automatically on first run.
    • Train the custom model.
    • Save model diagrams (if the diagram cell is executed).

7) Key Results / Performance

  • Not provided. The notebook demonstrates the mechanics of model subclassing and training rather than optimizing for maximum accuracy.

8) Screenshots / Sample Output

  • Not provided in repository. Running the diagram cell will save a .png visualization of the custom model locally.

9) Additional Learnings / Reflections

  • Subclassing tf.keras.Model offers complete freedom in architecture and training control.
  • This flexibility is useful for research, custom loss integration, and non-standard forward passes.
  • The project reinforces understanding of TensorFlow low-level operations combined with Keras high-level utilities.
  • Lower performance in some runs is intentional to focus on the API’s mechanics, not leaderboard scores.

👤 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.

اگر بخوای میتونم همین README رو هم با فرمت و نام‌گذاری هماهنگ با بقیه ریپازیتوری‌هات (مثل DL_TensorFlow_LowLevelAPI_CustomModel) بازنویسی کنم تا گیت‌هابت کاملاً یکدست بشه. آیا این کار رو انجام بدم؟