ResNet-18 Implementation from Scratch

A PyTorch implementation of the ResNet-18 architecture built without torchvision.models. This project was developed to deeply explore the mechanics and mathematics of deep residual learning, gradient flow, and weight initialization.

Performance

• Accuracy: 93.12%
• Dataset: CIFAR-10

Architecture Design

Implemented the core ResNet-18 components manually to learn the mathematical foundations of deep networks:

• Custom Residual Blocks: Built from scratch to handle identity mappings and skip connections.
• Skip Connections: Designed to mitigate the vanishing gradient problem, allowing for effective training of deeper layers.
• He Initialization: Used Kaiming Normal initialization to maintain variance across layers and stabilize the early stages of training.

Optimization Pipeline

• Optimizer: Stochastic Gradient Descent (SGD) with a momentum factor and L2 regularization (Weight Decay).
• Scheduling: Implemented a Learning Rate Scheduler to refine weights as the loss plateaued.
• Data Augmentation: Utilized Random Cropping and Horizontal Flipping to improve model generalization and prevent overfitting.

Skills & Tools

Languages: Python
Frameworks: PyTorch
Concepts: Computer Vision, Neural Networks, He Initialization, Data Augmentation, SGD Optimization