Function main

examples/dnn_semantic_segmentation_train_ex.cpp:159–291

// ----------------------------------------------------------------------------------------

int main(int argc, char** argv) try
{
    if (argc < 2 || argc > 3)
    {
        cout << "To run this program you need a copy of the PASCAL VOC2012 dataset." << endl;
        cout << endl;
        cout << "You call this program like this: " << endl;
        cout << "./dnn_semantic_segmentation_train_ex /path/to/VOC2012 [minibatch-size]" << endl;
        return 1;
    }

    cout << "\nSCANNING PASCAL VOC2012 DATASET\n" << endl;

    const auto listing = get_pascal_voc2012_train_listing(argv[1]);
    cout << "images in dataset: " << listing.size() << endl;
    if (listing.size() == 0)
    {
        cout << "Didn't find the VOC2012 dataset. " << endl;
        return 1;
    }

    // a mini-batch smaller than the default can be used with GPUs having less memory
    const unsigned int minibatch_size = argc == 3 ? std::stoi(argv[2]) : 23;
    cout << "mini-batch size: " << minibatch_size << endl;

    const double initial_learning_rate = 0.1;
    const double weight_decay = 0.0001;
    const double momentum = 0.9;

    bnet_type bnet;
    dnn_trainer<bnet_type> trainer(bnet,sgd(weight_decay, momentum));
    trainer.be_verbose();
    trainer.set_learning_rate(initial_learning_rate);
    trainer.set_synchronization_file("pascal_voc2012_trainer_state_file.dat", std::chrono::minutes(10));
    // This threshold is probably excessively large.
    trainer.set_iterations_without_progress_threshold(5000);
    // Since the progress threshold is so large might as well set the batch normalization
    // stats window to something big too.
    set_all_bn_running_stats_window_sizes(bnet, 1000);

    // Output training parameters.
    cout << endl << trainer << endl;

    std::vector<matrix<rgb_pixel>> samples;
    std::vector<matrix<uint16_t>> labels;

    // Start a bunch of threads that read images from disk and pull out random crops.  It's
    // important to be sure to feed the GPU fast enough to keep it busy.  Using multiple
    // thread for this kind of data preparation helps us do that.  Each thread puts the
    // crops into the data queue.
    dlib::pipe<training_sample> data(200);
    auto f = [&data, &listing](time_t seed)
    {
        dlib::rand rnd(time(0)+seed);
        matrix<rgb_pixel> input_image;
        matrix<rgb_pixel> rgb_label_image;
        matrix<uint16_t> index_label_image;
        training_sample temp;