model.md

Models

Models in tom use the model structure. A simple network can be made like so:

struct model m;

int main(void) {
    model_init(&m, 100);
    model_add_layer(&m, LAYER_DENSE, 10, 200);
    model_add_layer(&m, LAYER_RELU, 200, 200);
    model_set_loss(&m, LOSS_MSE);
    model_finalize(&m);

    model_free(&m);
}

model

The model object.

struct model {
    // First and last layers.
    struct layer *first, *last;

    // Number of layers.
    int n_layers;

    // Number of samples.
    int n_samples;

    // Loss object.
    struct loss loss;

    // We store the input, output and y matrices.
    struct matrix *input, *output, *y;

    // Loss output.
    struct matrix *loss_output;

    // Store the last gradient.
    struct matrix *last_gradient;
};

int model_init(struct model *obj, int n_samples)

Initialize an empty model object.

int model_free(struct model *obj)

Free a model object. Free all the layers, optimizers, and matrices, along with the loss.

struct layer* model_add_layer(struct model *obj, enum layer_type type, int input_size, int output_size)

Add and initialize a layer on the model. Returns the layer if successful.

struct layer* model_add_conv2d_layer(struct model* obj, int input_channels, int input_height, int input_width, int n_filters, int filter_size, int stride)

Add a conv 2D layer without initializing it. Returns the layer if successful.

struct layer* model_add_maxpool2d_layer(struct model* obj, int input_channels, int input_height, int input_width, int pool_size, int stride)

Add a max pooling 2D layer without initializing it. Returns the layer if successful.

struct layer* model_add_padding2d_layer(struct model* obj, int input_channels, int input_height, int input_width, int padding_x, int padding_y)

Add a padding 2D layer without initializing it. Returns the layer if successful.

void model_set_loss(struct model *obj, enum loss_type type)

Set the model's loss.

int model_finalize(struct model *obj)

Finalize and initialize the model.

int model_init_optimizers(struct model *obj, enum optimizer_type type, ...)

Initialize optimizers on the model.

int model_predict(struct model* obj, struct matrix* X, struct matrix* Y)

Predict. Takes an input and output matrix with any number of samples.

double model_calc_loss(struct model* obj, struct matrix* X, struct matrix* Y)

Calculate model loss.

int model_train(struct model* obj, struct matrix* X, struct matrix* Y, int epochs, bool debug)

Train the model.

int model_forward(struct model *obj, bool training)

Perform a forward pass on the model.

int model_backward(struct model *obj)

Perform a backward pass on the model.

int model_update(struct model* obj)

Update each trainable layer in the model.