Git Link --> https://github.com/prajjwalmehta123/Lenet5

LeNet-5 Implementation in C++

This repository contains a modern C++ implementation of the LeNet-5 convolutional neural network architecture, designed for MNIST digit classification. The implementation features both CPU (OpenMP) and GPU (CUDA) acceleration options.

Features

• Complete LeNet-5 architecture implementation
• MNIST dataset support
• OpenMP parallel processing for CPU acceleration
• Optional CUDA support for GPU acceleration
• Batch processing capability
• Adam optimizer implementation
• Modular design with separate layer implementations

Prerequisites

Required:

• C++17 compatible compiler
• CMake (minimum version 3.16)
• OpenMP

Optional:

• CUDA Toolkit (for GPU acceleration)
• Compatible NVIDIA GPU

Building the Project

1. Clone the repository:

git clone [repository-url]
cd Lenet5

2. Create a build directory:

mkdir build
cd build

3. Configure with CMake:

For CPU-only build:

cmake ..

For GPU-enabled build:

cmake -DUSE_CUDA=ON ..

4. Build the project:

make

Usage

1. Set up environment variables for MNIST dataset paths:

export MNIST_IMAGES_PATH=/path/to/train-images-idx3-ubyte
export MNIST_LABELS_PATH=/path/to/train-labels-idx1-ubyte
export MNIST_TEST_IMAGES_PATH=/path/to/t10k-images-idx3-ubyte
export MNIST_TEST_LABELS_PATH=/path/to/t10k-labels-idx1-ubyte

2. Run the executable:

./lenet5

Dataset

The implementation uses the MNIST dataset. You can download it from here.

The dataset should be in the IDX format:

• Training images: train-images-idx3-ubyte
• Training labels: train-labels-idx1-ubyte
• Test images: t10k-images-idx3-ubyte
• Test labels: t10k-labels-idx1-ubyte

Implementation Details

Network Architecture

• Input Layer: 32x32 grayscale images (padded from 28x28)
• C1: Convolutional layer (6 feature maps, 5x5 kernels)
• S2: Average pooling layer (2x2)
• C3: Convolutional layer (16 feature maps, 5x5 kernels)
• S4: Average pooling layer (2x2)
• F5: Fully connected layer (120 neurons)
• F6: Fully connected layer (84 neurons)
• Output: Fully connected layer (10 neurons)

Optimization

• Adam optimizer with configurable parameters
• ReLU activation functions
• OpenMP parallelization for CPU
• Optional CUDA acceleration for GPU

Performance

• CPU Version: Utilizes OpenMP for parallel processing
• GPU Version: Supports CUDA acceleration (requires -DUSE_CUDA=ON during build)
• Batch sizes:
  • CPU: Default batch size of 64
  • GPU: Default batch size of 256