Convolutional AutoEncoder for CIFAR Dataset

This project implements a Convolutional AutoEncoder model for image compression using the CIFAR-10 dataset. The autoencoder is a type of unsupervised neural network that learns to compress (encode) images into a latent representation and then reconstructs (decodes) the images to their original form.

Project Structure

project/
├── src/
│   ├── data_loader.py    # Data loading functions for CIFAR-10/100
│   ├── models.py         # Autoencoder model architectures
│   ├── utils.py          # Utility functions (visualization, PSNR calculation)
│   └── train.py          # Training script
├── notebooks/
│   └── Compression.ipynb # Original Jupyter notebook
├── data/                 # Dataset storage (auto-downloaded)
├── models/               # Trained models save here
├── logs/                 # Training logs save here
├── requirements.txt      # Python dependencies
└── README.md

Key Features

• Image Compression: Reduces image dimensionality while preserving key features
• Multiple Architectures: Basic and advanced autoencoder models
• CIFAR-10/100 Support: Works with both CIFAR datasets
• Performance Metrics: PSNR calculation for reconstruction quality
• Visualization Tools: Compare original vs reconstructed images

Quick Start

1. Install dependencies:

   pip install -r requirements.txt

2. Train the model:

   cd src
   python train.py

3. Or explore the original notebook:

   jupyter notebook notebooks/Compression.ipynb

Model Architectures

Basic Autoencoder (v1)

• Simple encoder-decoder structure
• 2 convolutional layers in encoder
• 2 transposed convolutional layers in decoder

Advanced Autoencoder (v2)

• 6 convolutional blocks with batch normalization
• Progressive downsampling: 32x32 → 1x1
• Dense latent representation
• Symmetric upsampling decoder

Performance Evaluation

• Loss Function: Binary crossentropy
• Optimizer: Adam
• Metrics: PSNR (Peak Signal-to-Noise Ratio)
• Callbacks: Early stopping, model checkpointing