Training Guide

How to train a neural network with Aicraft.

Overview

Aicraft provides a complete training pipeline: forward pass, loss computation, backward pass (autograd), and optimiser step.

Basic Training Loop

#include "aicraft/aicraft.h"

int main(void) {
    ac_init();

    // Define network
    AcLayer *net[] = {
        ac_dense(784, 128, AC_RELU),
        ac_dense(128, 10,  AC_SOFTMAX)
    };

    // Create optimiser
    AcOptimizer *opt = ac_adam(net, 2, 0.001f);

    // Training loop
    for (int epoch = 0; epoch < 10; epoch++) {
        float total_loss = 0;

        for (int i = 0; i < num_samples; i++) {
            ac_mem_checkpoint();

            AcTensor *x = get_input(i);
            AcTensor *target = get_label(i);

            // Forward
            AcTensor *pred = ac_forward_seq(net, 2, x);

            // Loss
            AcTensor *loss = ac_cross_entropy(pred, target);
            total_loss += ac_scalar(loss);

            // Backward
            ac_backward(loss);

            // Update weights
            ac_optimizer_step(opt);

            ac_mem_restore();
        }

        printf("Epoch %d — loss: %.4f\n", epoch, total_loss / num_samples);
    }

    ac_cleanup();
    return 0;
}

Optimisers

Optimiser	Function	Key Parameters
SGD	ac_sgd()	learning rate, momentum
Adam	ac_adam()	lr, beta1, beta2, epsilon
AdamW	ac_adamw()	lr, beta1, beta2, weight decay

Loss Functions

Loss	Function	Use Case
Cross-Entropy	ac_cross_entropy()	Classification
MSE	ac_mse()	Regression
Huber	ac_huber()	Robust regression

Tips

• Use ac_mem_checkpoint() / ac_mem_restore() inside the inner loop to keep memory constant
• Compile with -O3 and SIMD flags for maximum throughput
• Add -lvulkan to offload heavy operations to the GPU