Writing training routines for neural networks can be difficult, however, MagmaDNN provides the Model class, which only requires the coder to define the network and it will take care of training. The model we'll be using is magmadnn::model::NeuralNetwork<T>.
For a full neural network example using the MNIST data set see the simple_network example.
Neural Network models take four parameters on creation: a vector of layers, loss function, optimizer, and parameter struct. For example
/* initialize our model parameters */
model::nn_params_t params;
params.batch_size = 100; /* batch size: the number of samples to process in each mini-batch */
params.n_epochs = 5; /* # of epochs: the number of passes over the entire training set */
params.learning_rate = 0.05; /* learning rate of model */
auto input = layer::input(x_batch);
auto fc = layer::input(input->out(), 10);
auto act = layer::activation(fc->out(), layer::SOFTMAX);
auto output = layer::output(act->out());
std::vector<layer::Layer<float> *> layers = {input, fc, act, output}; /* create our vector of layers */
/* use cross entropy loss and stochastic gradient descent as an optimizer */
model::NeuralNetwork<float> model(layers, optimizer::CROSS_ENTROPY, optimizer::SGD, params);And that is it! Our model has been made and is ready for training. All Model<T> subclasses define a fit function, which will train on the data set. The fit function takes the data set, labels, metric struct, and verbosity options. For instance
model::metric_t metric_out;
bool is_verbose = true;
/* x_train and y_train are tensors containing the data set */
model.fit(x_train, y_train, labels, metric_out, is_verbose);
/* metric will now store the training accuracy, loss, and duration */