Self-Supervised Learning for Image Representation Learning

Overview

This project explores Self-Supervised Learning (SSL) through an image rotation prediction pretext task. The goal is to train a model to recognize image orientations, allowing it to learn meaningful representations from unlabeled data. These learned features are later transferred to a downstream image classification task.

Note: This project is a proof of concept designed to demonstrate the Self-Supervised Learning technique.

🔍 Project Details

Self-supervised learning eliminates the need for labeled datasets by leveraging pretext tasks that generate supervisory signals from raw data. In this project, we:

• Train a model to predict image rotations (0°, 90°, 180°, 270°) as the pretext task.
• Use MobileNetV2 as the feature extractor to improve learning efficiency.
• Fine-tune the model for a binary classification task (distinguishing between two classes: Duck and Fish).

📂 Dataset

• Dataset: Tiny ImageNet (subset of ImageNet).
• Pretext Task: Image rotation classification.
• Downstream Task: Binary classification with selected categories.

⚙️ Model Architecture

• Feature Extractor: MobileNetV2 (pretrained backbone).
• Data Augmentation: Horizontal flips, zoom, contrast adjustments.
• Classification Head: Dense layers with ReLU and softmax activation.
• Loss Function: Sparse categorical cross-entropy.

🚀 Implementation Details

• Frameworks: TensorFlow, Keras, NumPy, Matplotlib.
• Training Strategy:
  • Pretext Task: Train model on rotation prediction.
  • Fine-Tuning: Transfer learned features to the classification task.
  • Optimization: Adam optimizer, early stopping, learning rate scheduling.

📊 Results

• Training from scratch performed poorly due to unstable optimization.
• Using MobileNetV2 significantly improved performance.
• The fine-tuned model achieved high accuracy in the downstream task.

🖥️ Training Environment

• Platform: Google Colab.
• Hardware: NVIDIA Tesla T4 GPU.
• Python Version: 3.9.

Project Structure

Projet DL/
├── models/                 # Saved model checkpoints
├── src/                    # Source code
│   ├── data/               # Data processing scripts
│   ├── models/             # Model architecture definitions
│   └── utils/              # Utility functions
├── tests/                  # Unit tests
├── requirements.txt        # Python dependencies
└── README.md               # This file

Installation

Prerequisites

• Python 3.7+
• PyTorch 1.7+
• CUDA (for GPU acceleration, optional)

Setup

1. Clone the repository:

git clone https://github.com/haithem-ss/Self-supervised-learning
cd "Projet DL"

2. Create and activate a virtual environment (optional but recommended):

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

References

[1] YouTube, "Self-Supervised Learning Overview," available at: https://www.youtube.com/watch?v=CG9xbAfq6wI.

[2] Neptune.ai, "Self-Supervised Learning: What It Is and How It Works," available at: https://neptune.ai/blog/self-supervised-learning#:~:text=Self%2Dsupervised%20learning%20is%20a,as%20predictive%20or%20pretext%20learning.

[3] V7 Labs, "The Ultimate Guide to Self-Supervised Learning," available at: https://www.v7labs.com/blog/self-supervised-learning-guide.

[4] Shelf.io, "Self-Supervised Learning Harnesses the Power of Unlabeled Data," available at: https://shelf.io/blog/self-supervised-learning-harnesses-the-power-of-unlabeled-data/.

[5] Kaggle, "Tiny ImageNet," available at: https://www.kaggle.com/datasets/akash2sharma/tiny-imagenet.

[6] Tsang, S., "Review: SimCLR – A Simple Framework for Contrastive Learning of Visual Representations," Medium, available at: https://sh-tsang.medium.com/5de42ba0bc66.

[7] AI Multiple, "Self-Supervised Learning," available at: https://research.aimultiple.com/self-supervised-learning/#what-are-its-limitations.

More Resources

• Checkout my article about this project here.