SCENN Documentation

# Model

# Sequential Network

SequentialNetwork is a sequential neural network class in scenn namespace.

# constructor

SCENN_CONSTEXPR SequentialNetwork(loss_function, layers...)

loss_function is a loss function like MSE, CrossEntropy and so on. In fact, these loss functions are not functions, but structs. See the development guide for detail. layers are DenseLayers or ActivationLayers.

# example

SCENN_CONSTEXPR auto model = SequentialNetwork(CrossEntropy(), DenseLayer<10, 3, double>(), ActivationLayer<3, double>(Sigmoid())

In the above exapmle, model is a neural network with a cross entropy loss function, 10-dim input vectors and 3-dim output vectors.

# train

SCENN_CONSTEXPR auto train<MiniBatchSize>(training_data, epochs, learning_rate) => trained_model

# example

// model is an instance of SequentialNetwork class
SCENN_CONSTEXPR auto trained_model = model.train<2>(training_data, 10, 0.1)

# single_forward

SCENN_CONSTEXPR auto single_forward(test_data) => prediction

# evaluate

SCENN_CONSTEXPR auto evaluate(test_data) => number_of_accurate_prediction

# Detail

namespace scenn {
template <class LossFunction, class... Layers>
class SequentialNetwork {
  SCENN_CONSTEXPR SequentialNetwork(LossFunction &&loss, Layers &&... layers);
  SCENN_CONSTEXPR SequentialNetwork(const LossFunction &loss,
                                    const Layers &... layers);
  template <std::size_t MiniBatchSize, class Train, class T>
  SCENN_CONSTEXPR auto train(Train &&training_data, std::size_t epochs,
                             T &&learning_rate);
  template <class Test>
  SCENN_CONSTEXPR auto single_forward(Test &&x) const;
  template <class Test>
  SCENN_CONSTEXPR auto evaluate(Test &&test_data) const;
};
}