WandB-CPP

Installation

wandb-cpp works by wrapping the Weight & Biases.So, you need to install wandb according to this procedure first.

The required Python packages are listed in the requirements.txt file.

Usage

Basic Usage

#include <cmath>
#include <iostream>

#include "wandbcpp.hpp"

int main() {
  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"basic"}});

  int N = 100;

  wandbcpp::add_config({{"N", N}, {"mode", "abc"}});

  for (int i = 0; i < N; i++) {
    double t = M_PI * i / N;
    double x = std::sin(M_PI * i / N);
    double y = std::cos(M_PI * i / N);
    wandbcpp::log({{"t", t}, {"x", x}, {"y", y}});
    std::cout << "i : " << i << std::endl;
  }

  wandbcpp::finish();
}

result

Save File

#include "wandbcpp.hpp"

int main() {
  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"save file"}});
  wandbcpp::save(__FILE__);  // save this source code
  wandbcpp::finish();
}

result

wandb.Table

use add_data

#include <cmath>

#include "wandbcpp.hpp"

int main() {
  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"table"}});
  int I = 10;
  int J = 100;

  for (int i = 0; i < I; i++) {
    wandbcpp::Table table({"x", "y"});
    for (int j = 0; j < J; j++) {
      double x = (i + 1) * std::sin(M_PI * j / J);
      double y = (i + 1) * std::cos(M_PI * j / J);
      table.add_data({x, y});
    }
    wandbcpp::log({{"mytable", table}});
  }
  wandbcpp::finish();
}

result

use add_column

#include "wandbcpp.hpp"

int main() {
  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"table"}});
  wandbcpp::Table table;
  std::vector<double> x = {1.0, 2.0, 3.0};
  std::vector<int> y = {4, 5, 6};
  table.add_column("x", {x.begin(), x.end()});
  table.add_column("y", {y.begin(), y.end()});
  wandbcpp::log({{"table", table}});
  wandbcpp::finish();
}

result

passing data in the constructor

#include "wandbcpp.hpp"

int main() {
  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"table"}});
  int I = 10;
  for (int i = 0; i < I; i++) {
    wandbcpp::Table table({"x", "y"}, {{"poyo", i}, {"hoge", i + 1}});
    wandbcpp::log({{"table", table}});
  }
  wandbcpp::finish();
}

result

wandb.Object3D

plot pointcloud

#include <array>
#include <random>

#include "wandbcpp.hpp"

std::array<double, 3> gen_random_point() {
  static std::random_device seed_gen;
  static std::mt19937 engine(seed_gen());
  static std::uniform_real_distribution<> dist(-1.0, 1.0);
  return {dist(engine), dist(engine), dist(engine)};
}

int main(int argc, char const* argv[]) {
  namespace np = wandbcpp::numpy;

  wandbcpp::init({.project = "example_wandb_cpp", .tags = {"object3d"}});

  int num_points = 300;

  std::vector<std::array<double, 3>> point_cloud(num_points);

  std::generate(point_cloud.begin(), point_cloud.end(), &gen_random_point);

  auto lst = wandbcpp::topylist(point_cloud.begin(), point_cloud.end());

  wandbcpp::log({{"obj3d", wandbcpp::Object3D{np::ndarray{lst}}}});

  wandbcpp::finish();

  return 0;
}

result

Build examples

mkdir build && cd build
cmake -D BUILD_WANDBCPP_EXE=ON ..
make

Implementation & Performance

This library executes most of its operations in multi-threaded mode. Therefore, this library should have little impact on the performance of the main processing.

Thanduriel / wandb-cpp