Cuda-ml is a C++ machine learning framework with automatic differentiation and cuda acceleration. It is built completely from scratch with a custom tensor class and custom cuda kernels. Dynamic and eager execution are key properties of its tapeless automatic differentiation. So the function is defined simply by running it and no forward graph is built.

Installation

cmake . -B build
cmake --build build
sudo make install -C build

Learning an Image

Coordinates → Color

#include <cuda-ml/cuda-ml.h>

Tensor targets{};
Tensor coordinates{};
int height{};
int width{};
read_image("image.png", targets, height, width);
create_coordinates(height, width, coordinates);
normalize({255}, targets);
normalize({static_cast<float>(height - 1), static_cast<float>(width - 1)}, coordinates);

MultiLayerPerceptron network{2, {64, 1}};
StochasticGradientDescent optimizer{network.parameters(), 0.001};

const size_t n_epochs{ 5000 };
const size_t print_epochs{ 100 };
for (int epoch = 0; epoch < n_epochs; ++epoch) {
    const Tensor predictions{ network(coordinates) };
    const Tensor loss{ mean_squared_error(predictions, targets) };
    loss.backward();
    optimizer.step();
    optimizer.zero_gradients();
    if ((epoch + 1) % print_epochs == 0)
        std::cout << "epoch " << epoch + 1 << " loss " << loss[{0, 0}] << '\n';
}

network.detach();
Tensor predictions{ network(coordinates) };
normalize({1. / 255.}, predictions);
write_image("reconstruction.png", predictions, height, width);

Execution

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib
nvcc -o learn_image learn_image.cu -lpng -lcuda-ml
./learn_image

Tensor Class

Construction

Tensor::from_scalar(1, {2, 2})
Tensor::from_vector({1, 2, 3, 4}, {2, 2})
Tensor::random_uniform(0, 1, {2, 2})
Tensor::random_normal(0, 1, {2, 2})

Operations

tensor1 + tensor2
tensor1 - tensor2
tensor1 * tensor2
tensor1 / tensor2
mm(tensor1, tensor2)
square(tensor)
relu(tensor)
sum(tensor)

Indexing

tensor[{1, 2}]

Automatic Differentiation

tensor1.requires_gradients();
product = tensor1 * tensor2;
product.backward();
gradients = tensor1.gradients();
tensor1.detach();

Output

std::cout << Tensor::from_vector({1, 2, 3, 4}, {2, 2});