Implementation of a Convolutional Neural Network Using CuPy

Table of Contents

• Overview
• About CuPy
• Objectives
• Methodology
• Results
• Conclusion
• Documentation

---

Overview

This project presents a complete implementation of a Convolutional Neural Network (CNN) from first principles using the CuPy library, enabling efficient GPU-accelerated computation in Python. It aims to explore the internal workings of CNNs beyond the abstraction provided by widely used frameworks such as TensorFlow or PyTorch. You can read

---

About CuPy

CuPy is an open-source array library for GPU-accelerated computing with Python. It offers a NumPy-compatible interface and leverages NVIDIA CUDA libraries such as cuBLAS, cuDNN, and cuFFT to perform high-speed numerical operations on the GPU. CuPy enables seamless transition from CPU-based NumPy code to GPU execution, making it ideal for scientific computing and deep learning research.

---

Objectives

• Develop essential CNN components (e.g., convolutional layers, pooling, activation functions, fully connected layers, dropout, and loss functions) as modular Python classes
• Employ optimization strategies like im2col, col2im, and memory-efficient techniques such as as_strided for convolution operations
• Construct a training pipeline using a custom Sequential model and implement the AdamOptimizer for gradient-based learning
• Validate the model using the MNIST dataset by assessing classification accuracy, loss, and training time across various batch sizes

---

Methodology

• Preprocess MNIST images: convert to grayscale .png format and normalize input values
• Implement forward and backward propagation without loops, leveraging matrix operations and GPU acceleration
• Analyze model performance through validation curves and confusion matrices

---

Conclusion

This implementation confirms the feasibility of building a performant CNN purely with CuPy, highlighting both computational efficiency and the educational value of low-level deep learning architectures. It provides a foundation for extending the model to more complex datasets and deeper network designs.---

---

Documentation

A full project report detailing methodology, implementation, results, and analysis is available in the DOCS folder:

Project Report (PDF)

---

Track the progress of the project here: GitHub Project Board

---