Download the PHP package numpower/autograd without Composer

NumPower Autograd

NumPower Autograd enriches the NumPower extension by enabling automatic differentiation through reverse-mode (backpropagation) automatic differentiation. Automatic differentiation is a computational technique used to efficiently and accurately compute derivatives of functions. Unlike numerical differentiation, which can be prone to errors and inefficiencies, automatic differentiation systematically applies the chain rule to compute gradients.

Reverse-mode automatic differentiation, commonly known as backpropagation, is particularly powerful in machine learning and optimization tasks. It calculates gradients of scalar functions with respect to N-dimensional arguments in a highly efficient manner, making it ideal for training complex models.

NumPower Autograd also supports GPU acceleration, leveraging the power of graphics processing units for faster computation. This capability enhances the library's performance, particularly for large-scale mathematical and machine learning tasks.

By integrating backpropagation and GPU support, NumPower Autograd provides high-performance gradient computation capabilities, facilitating advanced mathematical and machine learning applications.

Requirements

• NumPower Extension >= 0.5.x (https://github.com/NumPower/numpower)
• PHP >= 8.3
• Composer

Installing

Getting Started

The NumPower\Tensor class is a type of N-dimensional array with support for automatic differentiation. It uses the NDArray object from the NumPower extension as its engine, retaining many of the flexibilities of using an NDArray, such as using it as an operator in arimetic operations, simplifying code reading.

This code demonstrates a simple example of automatic differentiation using variables and operations. It computes the gradients of the result of a computation ($c) with respect to its inputs ($a and $b).

Using a video card with CUDA support

If you have compiled the NumPower extension with GPU utilization capabilities, you can perform operations on the GPU by allocating your Tensor in VRAM. The operations will automatically identify whether your Tensor is in RAM or VRAM and will perform the appropriate operation on the correct device.

Tensor works exactly the same as NDArray when dealing with different devices in multi-argument operations. Unless the argument is a scalar, all N-dimensional arguments of an operation must be stored on the same device.

Simple training with autograd

Here we can see a more practical example. Let's create a neural network with a hidden layer and 1 output layer. Some common Neural Network functions are ready-made and can be accessed statically through the NumPower\NeuralNetwork module.

If you want to know more about each step of creating this model, access our documentation and see step-by-step instructions on how to build this model.

The above model can be used for several different classification problems. For simplicity, let's see if our model can solve the XOR problem.