SO, Usul-cpp is a library I've written in C++, it is based around allowing data scientists to create their models without heavy libraries such as tensorflow, and just express their models using maths, and Usul will handle the rest. There is builtin tensor support, auto-grad, back-prop and its quite fast.
I wrote it because, it seemed less mystifying to read maths instead of a random class that supposedly does something. I wrote it from scratch in cpp (23) due to the fact that I wanted to emphasise on precision, however it's subtly built on top of xtensor.
check it out in action #include You can view the project at [https://github.com/crushr3sist/Usul-cpp](https://github.com/crushr3sist/Usul-cpp) Any ideas for what to do next are all welcome, you can also pull and implement features that you think could improve this project. (no vibecoded prs please, this project is not vibe-coded) Comments URL: https://news.ycombinator.com/item?id=50440232 Points: 2 # Comments: 1
which will output int main() { double LR = 0.1; int EPOCH = 5000; // lets write the forward pass Tensor X = { {0.0, 0.0}, // this is what we're feeding it {0.0, 1.0}, {1.0, 0.0}, {1.0, 1.0}, }; // this is what we're after Tensor Y_true = { 0.0, // xor{0.0, 0.0} = 0.0 1.0, // xor{0.0, 1.0} = 1.0 1.0, // xor{1.0, 0.0} = 1.0 0.0 // xor{1.0, 1.0} = 0.0 }; Y_true.reshape({4, 1}); // we are now going to create a single layer, double neuron network // each layer will have a bias and weights tensor. auto W1 = generate_random_matrix({2, 4}, -1.0, 1.0); auto B1 = generate_random_matrix({1, 4}, -1.0, 1.0); auto W2 = generate_random_matrix({4, 1}, -1.0, 1.0); auto B2 = generate_random_matrix({1, 1}, -1.0, 1.0); // this is the training loop for (int i = 0; i < EPOCH; i++) { // so lets do one forward pass // first layer auto Z1 = X % W1 + B1; auto A = Z1.ReLU(); // second layer auto Z2 = A % W2 + B2; // activation // should be - auto S = Z2.activate(Activate::Sigmoid); auto S = Z2.Sigmoid(); // Loss // should be - auto Loss = S.loss(Loss::MSE); auto loss = (S - Y_true) * (S - Y_true); loss = loss.mean(); cout << "Loss: " << loss << "\n"; cout << "guesses: " << "\n" << S << "\n"; cout << "should be: " << Y_true << "\n"; loss.backwards(); W1.step(LR); B1.step(LR); W2.step(LR); B2.step(LR); } }
Just write the maths, and Usul-cpp will handle the rest, it is extremely lightweight, and the entire library is memory safe. And its called Usul, because I'm a huge dune fan. Guesses: Tensor({{ 0.027546}, { 0.963592}, { 0.964138}, { 0.038964}}, shape=(4,1)) should be: Tensor({{ 0.}, { 1.}, { 1.}, { 0.}}, shape=(4,1)) Loss: Tensor( 0.001222, shape=())
