Computer Vision Architectures – From Scratch Implementations

Overview

This repository is a structured exploration of classical and modern computer vision architectures implemented in PyTorch.

The goal of this project is to deeply understand the evolution of convolutional and attention-based models — from AlexNet to Vision Transformers — by implementing them modularly and analyzing their architectural trade-offs.

---

Architectures Implemented

• AlexNet
• VGG
• ResNet
• DenseNet
• GoogLeNet (Inception v1)
• MobileNet
• SqueezeNet
• Vision Transformer (ViT)

---

Additional Modules

• Convolution operations (from scratch implementation)
• Regularization techniques
• Transfer learning strategies

---

Research Motivation

This repository was created to:

• Understand architectural innovations in deep learning
• Analyze parameter efficiency vs performance trade-offs
• Compare convolution-based and attention-based approaches
• Explore generalization techniques in deep neural networks

---

Key Insights from Implementation

• Residual connections mitigate vanishing gradients
• Dense connectivity encourages feature reuse
• Depthwise separable convolutions reduce computational cost
• Transformers remove spatial locality bias but require larger datasets
• Regularization techniques significantly improve generalization

---

Tech Stack

• Python
• PyTorch
• NumPy

---

Future Work

• Hybrid CNN-Transformer architectures
• 3D CNNs for hyperspectral imagery
• Vision-Language Models (VLMs)
• Self-Supervised Pretraining methods

---

This repository was created as a structured study of deep learning architectures to understand their mathematical foundations and architectural evolution.