MNIST

Training a neural network to recognize hand-written digits from MNIST.

Files taken from https://github.com/Jaswar/nnlib/tree/main/examples/mnist.

The main.cpp file expects one argument which is the absolute path to the MNIST_train.txt file. The file can be downloaded from https://github.com/halimb/MNIST-txt.

 
#include <iostream>
#include <nnlib/network.h>
#include <nnlib/read.h>
#include <nnlib/verify.cuh>
#include <nnlib/onehot_encode.h>
 
int main(int argc, char** argv) {
    if (argc < 2) {
        std::cout << "Dataset file path was not specified." << std::endl;
        return 1;
    }
 
    showCudaInfo();
 
    sTensor dataset = readCSV(argv[1], ",", 4);
    sTensor X = std::make_shared<Tensor>(dataset->shape[0], dataset->shape[1] - 1);
    sTensor yv = std::make_shared<Tensor>(dataset->shape[0]);
 
    for (int i = 0; i < dataset->shape[0]; i++) {
        yv->data[i] = dataset->data[i * dataset->shape[1] + 0];
        for (int j = 1; j < dataset->shape[1]; j++) {
            X->data[i * X->shape[1] + j - 1] = dataset->data[i * dataset->shape[1] + j] / 255;
        }
    }
 
    sTensor y = oneHotEncode(yv);
 
    std::cout << y << std::endl;
 
    Network network = Network(X->shape[1], true);
    network.add(64);
    network.add(y->shape[1], "sigmoid");
 
    std::vector<Metric*> metrics = {new CategoricalAccuracy(), new MeanSquaredError()};
    network.train(X, y, 10, 10, 0.01, new CategoricalCrossEntropy(), metrics);
 
    return 0;
}

The project can be built with the following CMake script. This script requires CMAKE_PREFIX_PATH to be set to the install directory of nnlib.

cmake_minimum_required(VERSION 3.20)
include(CheckLanguage)
 
project(mnist_nnlib LANGUAGES CXX)
set(CMAKE_CXX_STANDARD 14)
 
find_package(nnlib CONFIG REQUIRED)
 
check_language(CUDA)
if (CMAKE_CUDA_COMPILER)
    enable_language(CUDA)
    set(CMAKE_CUDA_STANDARD 14)
 
    include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
 
    # Add flag to specify the architecture of the GPU (compute capability)
    if (NOT DEFINED CUDA_ARCHITECTURES)
        set_target_properties(nnlib PROPERTIES CUDA_ARCHITECTURES "50")
    endif()
 
    set_target_properties(nnlib PROPERTIES
            CUDA_SEPARABLE_COMPILATION ON
            CUDA_RESOLVE_DEVICE_SYMBOLS ON)
endif()
 
add_executable(${PROJECT_NAME} main.cpp)
 
target_link_libraries(${PROJECT_NAME} PRIVATE nnlib)