CapsNet Vision Transformer Hybrid

A comparative study of Capsule Networks, Vision Transformers, and hybrid architectures for CIFAR-100 image classification.

Overview

This repository contains implementations and experiments comparing different deep learning architectures:

• CNN Baseline: Standard convolutional neural network baseline
• Vision Transformer (ViT): Transformer-based image classification model
• CapsNet: Capsule Network implementation with dynamic routing
• CapsViT Hybrid: Novel hybrid architecture combining Capsule Networks with Vision Transformer features

Project Structure

• 0. CNN_Baseline/ - Baseline CNN implementation
• 1. vit_run/ - Vision Transformer training code and results
• 2. CapsNet/ - Capsule Network baseline implementation
• 3. capsvit/ - Hybrid CapsNet-ViT architecture
• data/ - CIFAR-100 dataset
• Evaluation/ - Model evaluation scripts and metrics

Requirements

• PyTorch
• torchvision
• einops
• GPUtil
• numpy

Usage

Each model can be trained independently using the Python scripts in their respective directories. Models are trained on CIFAR-100 dataset with 100 classes.

Results

Results and trained models are stored in the respective output directories for each architecture.