CUDANet/examples/alexnet/alexnet.cpp

#include <cudanet.cuh>
#include <opencv2/opencv.hpp>
#include <string>
#include <vector>

std::vector<float>
readAndNormalizeImage(const std::string &imagePath, int width, int height) {
    // Read the image using OpenCV
    cv::Mat image = cv::imread(imagePath, cv::IMREAD_COLOR);

    // Resize and normalize the image
    cv::resize(image, image, cv::Size(width, height));
    image.convertTo(image, CV_32FC3, 1.0 / 255.0);

    // Normalize the image https://pytorch.org/hub/pytorch_vision_alexnet/
    cv::Mat mean(image.size(), CV_32FC3, cv::Scalar(0.485, 0.456, 0.406));
    cv::Mat std(image.size(), CV_32FC3, cv::Scalar(0.229, 0.224, 0.225));
    cv::subtract(image, mean, image);
    cv::divide(image, std, image);

    // Convert the 3D image matrix to a 1D array of floats
    std::vector<float> imageData;
    for (int c = 0; c < image.channels(); ++c) {
        for (int i = 0; i < image.rows; ++i) {
            for (int j = 0; j < image.cols; ++j) {
                imageData.push_back(image.at<cv::Vec3f>(i, j)[c]);
            }
        }
    }

    return imageData;
}

CUDANet::Model *createModel(
    const shape2d inputSize,
    const int inputChannels,
    const int outputSize
) {
    CUDANet::Model *model =
        new CUDANet::Model(inputSize, inputChannels, outputSize);

    // Block 1
    CUDANet::Layers::Conv2d *conv1 = new CUDANet::Layers::Conv2d(
        inputSize, inputChannels, {11, 11}, {4, 4}, 64, {2, 2},
        CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("features.0", conv1);  // Match pytorch naming
    CUDANet::Layers::MaxPooling2d *pool1 = new CUDANet::Layers::MaxPooling2d(
        {56, 56}, 64, {3, 3}, {2, 2}, {0, 0}, CUDANet::Layers::ActivationType::NONE
    );
    model->addLayer("pool1", pool1);

    // Block 2
    CUDANet::Layers::Conv2d *conv2 = new CUDANet::Layers::Conv2d(
        {27, 27}, 64, {5, 5}, {1, 1}, 192, {2, 2}, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("features.3", conv2);
    CUDANet::Layers::MaxPooling2d *pool2 = new CUDANet::Layers::MaxPooling2d(
        {27, 27}, 192, {3, 3}, {2, 2}, {0, 0}, CUDANet::Layers::ActivationType::NONE
    );
    model->addLayer("pool2", pool2);

    // Block 3
    CUDANet::Layers::Conv2d *conv3 = new CUDANet::Layers::Conv2d(
        {13, 13}, 192, {3, 3}, {1, 1}, 384, {1, 1}, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("features.6", conv3);

    // Block 4
    CUDANet::Layers::Conv2d *conv4 = new CUDANet::Layers::Conv2d(
        {13, 13}, 384, {3, 3}, {1, 1}, 256, {1, 1}, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("features.8", conv4);

    // Block 5
    CUDANet::Layers::Conv2d *conv5 = new CUDANet::Layers::Conv2d(
        {13, 13}, 256, {3, 3}, {1, 1}, 256, {1, 1}, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("features.10", conv5);
    CUDANet::Layers::MaxPooling2d *pool5 = new CUDANet::Layers::MaxPooling2d(
        {13, 13}, 256, {3, 3}, {2, 2}, {0, 0}, CUDANet::Layers::ActivationType::NONE
    );
    model->addLayer("pool5", pool5);

    // Classifier
    CUDANet::Layers::Dense *dense1 = new CUDANet::Layers::Dense(
        6 * 6 * 256, 4096, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("classifier.1", dense1);

    CUDANet::Layers::Dense *dense2 = new CUDANet::Layers::Dense(
        4096, 4096, CUDANet::Layers::ActivationType::RELU
    );
    model->addLayer("classifier.4", dense2);

    CUDANet::Layers::Dense *dense3 = new CUDANet::Layers::Dense(
        4096, 1000, CUDANet::Layers::ActivationType::NONE
    );
    model->addLayer("classifier.6", dense3);

    return model;
}

int main(int argc, const char *const argv[]) {
    if (argc != 3) {
        std::cerr << "Usage: " << argv[0] << "<model_weights_path> <image_path>"
                  << std::endl;
        return 1;  // Return error code indicating incorrect usage
    }

    // Path to the image file
    std::string modelWeightsPath = argv[1];
    std::string imagePath        = argv[2];

    const shape2d inputSize     = {227, 227};
    const int inputChannels = 3;
    const int outputSize    = 1000;

    CUDANet::Model *model = createModel(inputSize, inputChannels, outputSize);

    model->validate();

    model->loadWeights(modelWeightsPath);

    // Read and normalize the image
    std::vector<float> imageData =
        readAndNormalizeImage(imagePath, inputSize.first, inputSize.second);

    // Print the size of the image data
    const float *output = model->predict(imageData.data());

    // Get max index
    int maxIndex = 0;
    for (int i = 0; i < outputSize; i++) {
        if (output[i] > output[maxIndex]) {
            maxIndex = i;
        }
    }

    std::string classLabel = CUDANet::Utils::IMAGENET_CLASS_MAP.at(maxIndex);

    std::cout << "Prediction: " << maxIndex << " " << classLabel << std::endl;
    return 0;
}