Matcha

Tensor arithmetic | Automatic differentiation | JIT compilation | Dataset pipelines | Neural networks | Gotchas

What it is

Matcha is a framework for optimized tensor arithmetic and machine learning. It features a very intuitive interface inspired by Numpy and Keras. Matcha brings all this to C++. It also provides a way to accelerate itself by Just-In-Time inspecting and modifying the structure of given tensor functions and compiling them into a set of instructions. On top of that, Matcha delivers a seamless dataset pipeline system, automatic differentiation system, and neural networks framework.

#include <iostream>
#include <matcha>

using namespace matcha;

int main() {
  Net net {
    nn::flatten,                             // flatten the inputs
    nn::Fc{300, "relu,batchnorm"},           // hidden layer
    nn::Fc{10, "softmax"}                    // output layer
  };

  Dataset mnist = load("mnist_train.csv");   // load the MNIST dataset
  net.loss = nn::Nll{};                      // use the negative log likelihood loss
  net.fit(mnist.batch(64));                  // fit the model

  tensor digit = load("digit.png");          // load a single digit image
  tensor probabilities = net(digit);         // make a prediction
  tensor result = argmax(probabilities);     // voila

  std::cout << "it is " << result << " with "
            << probabilities[result] * 100 << "% probability" << std::endl;
}

For more, check out tutorials.

License

Matcha is open source. It is available under the MIT license. It may be freely used and distributed.

Plans (and what's currently missing)

• More thorough CMake integration
• Python interface, Java/Kotlin interface
• Differentiable and JIT-compilable conditions and loops
• GPU acceleration (OpenCL, then CUDA)

Troubleshooting

Please, open an issue under the appropriate package on GitHub.

Contribute

See you on GitHub!