A Torch-like C++ framework for building neural network graphs. Runs on an automatic differentiation tensor library built on top of vecLib (with plans to move to a more cross plaform framework). A small collection of prebuilt layers, loss functions and optimisers are avaiable.

Design the neural network object you want to use from the nn::Net base object. Create the forward feed function using the autodiff::var for variables and Parameter for parameters. Prebuilt layers with parameters are available.

class Network : public nn::Net{
public:
    Model() : nn::Net(), fc1(10, 5, this), fc2(5, 1, this){}
    
    nn::FullyConnected fc1, fc2;
    autodiff::Var forward(autodiff::Var& x){
        auto y = fc1(x);
        auto z = fc2(y);
        return z;
    }
};

After initialization the network is registered with an optimizer which has access to the parameters and their gradients.

Network net;
opt::GD Optim(net.params);

Training is performed using the forward(x) method, and backpropigation using backwardProp(l). Optimization is then performed using the step() method.

auto output = net.forward(x);
auto l = loss::SoftMax(output, targets);
net.backward(l);
opt.step();