Machine Learning in C++: Overview of Key Libraries

Machine Learning in C++: Overview of Key Libraries

Although Python has become the de facto standard for research and rapid prototyping in the field of machine learning (ML), C++ remains an indispensable language for production systems where performance, low memory consumption, and real-time operation are critical. C++ allows you to get the maximum out of hardware, which is especially important when deploying models on mobile devices, embedded systems, or high-load servers.

In this article, we will review five leading machine learning libraries for C++, their strengths, typical use cases, and provide practical code examples. You will learn which library to choose for classification, regression, clustering, working with neural networks, or signal processing.

1. TensorFlow C++ API — The Power of Deep Learning

TensorFlow is one of the most popular frameworks for deep learning. Its C++ API allows you to load pre-trained models (e.g., from Python) and perform inference at maximum speed. This is an ideal choice for embedded systems (Raspberry Pi, Jetson Nano, Android) and production servers.

Key features:

• Working with saved computation graphs (SavedModel, Frozen Graph).
• GPU support via CUDA and cuDNN.
• Model optimization for a specific device (XLA, Quantization).

Example of loading and running a model (inference):

#include <tensorflow/core/public/session.h>#include <tensorflow/core/platform/env.h>

using namespace tensorflow;

int main() {    // Create a session    Session* session;    NewSession(SessionOptions(), &session);

    // Load the graph from a file    GraphDef graph_def;    Status status = ReadBinaryProto(Env::Default(), "model.pb", &graph_def);    if (!status.ok()) { /* error handling */ }

    // Add the graph to the session    session->Create(graph_def);

    // Prepare input data (tensor)    Tensor input_tensor(DT_FLOAT, TensorShape({1, 224, 224, 3}));    auto input_map = input_tensor.tensor<float, 4>();    // ... fill the tensor with data ...

    // Run inference    std::vector<Tensor> outputs;    session->Run({{"input_tensor", input_tensor}}, {"output_tensor"}, {}, &outputs);

    // Use the result    auto result = outputs[0].tensor<float, 2>();    std::cout << "Prediction: " << result(0, 0) << std::endl;

    session->Close();    return 0;}

2. Dlib — A Versatile Tool for ML and Computer Vision

Dlib is a general-purpose library that includes both classical machine learning algorithms (SVM, decision trees, K-means) and modern deep neural networks. It is particularly strong in computer vision tasks: face detection, object tracking, face recognition (FaceNet).

Key features:

• Easy integration: header-only for most functions.
• Built-in support for multithreading (thread_pool).
• Ready-to-use utilities for image processing and matrix operations.

Example of training an SVM classifier:

#include <dlib/svm.h>#include <dlib/rand.h>

using namespace dlib;

int main() {    // Kernel type: Radial Basis Function (RBF)    typedef radial_basis_kernel<sample_type> kernel_type;

    // Create a classifier and trainer    svm_c_trainer<kernel_type> trainer;    trainer.set_kernel(kernel_type(0.1));

    // Training data    std::vector<sample_type> samples;    std::vector<double> labels;    // ... fill samples and labels ...

    // Training    decision_function<kernel_type> df = trainer.train(samples, labels);

    // Test a new sample