Efficient CIFAR-100 Classification and Image Generation

This project adapts the Residual Network (ResNet) architecture with depthwise separable convolutions for efficient CIFAR-100 classification, achieving performance under 100,000 parameters and training in 10,000 steps. It also introduces a modified conditional Spectral Normalization GAN (SN-GAN) for image generation on CIFAR-100, optimized to run with less than 1,000,000 parameters.

Key Features

• Adapted ResNet Architecture: Utilizes depthwise separable convolutions to enhance efficiency for CIFAR-100 classification tasks.
• Modified SN-GAN for Image Generation: Implements adaptive weighting in hinge loss calculations to improve performance.
• Conditional Batch Normalization and Projection Discriminator: Ensures stable training and targeted image synthesis.

Overview

This project focuses on two main aspects:

1. Efficient Classification: A modified ResNet architecture that reduces computational complexity while maintaining high accuracy.
2. Stable Image Generation: A novel SN-GAN approach tailored for CIFAR-100, integrating advanced normalization techniques and a projection discriminator for more accurate and consistent results.

Dataset

• CIFAR-100: Both the classification and image generation models are trained and evaluated on the CIFAR-100 dataset.

Objectives

• Improve the efficiency of deep learning models for image classification tasks.
• Develop robust generative models capable of producing high-quality images with targeted characteristics.

Methods

• Depthwise Separable Convolutions: Applied in the ResNet architecture to optimize computational cost.
• Adaptive Weighting in Hinge Loss: Enhances the GAN training process by dynamically adjusting loss contributions.
• Conditional Batch Normalization: Provides instance-specific scaling and shifting to stabilize GAN training.
• Projection Discriminator: Utilized in the SN-GAN to enforce more precise class-conditional image generation.

Results

Classifier Performance

The modified ResNet shows competitive performance on CIFAR-100 with reduced computational requirements. Upon completion, it achieved:

• Training Accuracy: 65.5% ± 2.6%
• Test Accuracy: 57.7% ± 2.3%

Generator Performance

The adapted SN-GAN demonstrates stable and high-quality image generation capabilities. The results measured by the LPIPS (Learned Perceptual Image Patch Similarity) metric yielded a score of 0.23.

Future Work

• Explore further optimizations in the ResNet architecture.
• Investigate alternative loss functions and regularization techniques for GANs.