This project implements the classic LeNet-5 architecture using raw PyTorch tensors. It features a custom-built CNN with hand-crafted convolutional layers, Sigmoid activations, and max pooling. The bespoke training system combines stochastic gradient descent with cross-entropy loss, complemented by an efficient data pipeline. Tested on the Fashion MNIST dataset, this implementation offers a deep dive into CNN fundamentals, balancing performance with educational insights.
The project utilizes the following technologies:
