Learn the core concepts of microgradpp — the Value class, automatic differentiation, and building your first neural network.

Basic Usage#

Alright, let's get to the fun part! Here's everything you need to know to start using microgradpp.

The Value Class#

The Value class is the heart of microgradpp. Every number in your computation is wrapped in a Value, and microgradpp tracks all the operations so it can compute gradients automatically.

C++

#include <iostream>
#include "Value.h"
 
int main() {
    // Create values
    auto a = microgradpp::Value::create(2.0);
    auto b = microgradpp::Value::create(3.0);
 
    // Build a computation graph
    auto c = a * b;
 
    // Run backpropagation
    c->backProp();
 
    std::cout << "a->grad: " << a->grad << std::endl;  // 3.0
    std::cout << "b->grad: " << b->grad << std::endl;  // 2.0
 
    return 0;
}

When you call backProp(), microgradpp walks backward through your computation graph and fills in the .grad field for every Value involved. That's reverse-mode autodiff in action!

More Operations#

You can chain operations and use powers too:

C++

auto x = microgradpp::Value::create(2.0);
auto y = microgradpp::Value::create(3.0);
 
// Chain multiple operations
auto z = x * y + x->pow(2);
z->backProp();
 
// x->grad and y->grad now hold the correct gradients

Building a Neural Network#

Once you're comfortable with Value, building a full neural network is just a few lines away. microgradpp's MLP class makes it super easy:

C++

#include "microgradpp/MLP.hpp"
 
// Create a Multi-Layer Perceptron
// Input: 3 features | Hidden layers: 4, 4 neurons | Output: 1 neuron
MLP model(3, {4, 4, 1});
 
// Forward pass
auto output = model(input_data);
 
// Training loop
model.zeroGrad();      // Clear gradients from the previous step
loss->backProp();      // Compute gradients
// Update weights...   // Adjust parameters

Always call model.zeroGrad() before backProp() in your training loop. Otherwise, gradients from previous iterations will accumulate and your training will go haywire!

Activation Functions#

microgradpp supports several activation functions out of the box:

ReLU — great for hidden layers
Tanh — smooth and bounded
Sigmoid — perfect for binary outputs

These can be configured when setting up your layers.