nanograd-bgriebel 0.1.2

Scalar Automatic Differentiation Engine

Nanograd

A small scalar valued automatic differentiation engine written in C++ with bindings in python. Also includes a small neural network api with basic classes like Neuron, Layer, MultiLayerPerceptron for creating neural networks. Based on micrograd from Andrej Karpathy.

Installation

Nanograd is available on pypi (the python package index), and so can be installed with uv or pip. It's recommended to create a virtual environment when installing.

With pip:

# Create a virtual environment
python -m venv .venv # Creates a virtual environment called .venv
# Activate (Linux)
source .venv/bin/activate
# Activate (Windows)
.\.venv\env\scripts\activate
# Then you can install using pip
pip install nanograd-bgriebel

With uv:

# Create a virtual environment
uv venv
# Activate (Linux)
source .venv/bin/activate
# Activate (Windows)
.\.venv\env\scripts\activate
# Install using uv
uv pip install nanograd-bgriebel

Usage

Nanograd is built around the Value class, which is essentially a wrapped float which keeps track of calculations it is used in so that the gradients can be calculated.

Creating a Value is easy, it just needs a float as input:

import nanograd_bgriebel as ng
my_value = ng.Value(1.0)

Then Values can be combined with simple arithmetic operations, and gradients can be calculated.

import nanograd_bgriebel as ng
# Create some Values
x = ng.Value(2.0)
y = ng.Value(3.0)
z = ng.Value(4.0)

# Can add, and multiply the different Values
a = x+y
b = x*z

# And also raise them to different powers
c = x**3

# This can be chained to build up more complex expressions
d = a/(b*c) # equivalent to (x+y) / (x*z*(x**3))

# Then the result of the calculations can be obtained through the data property
print(d.data) # Output: 0.078125

# And the value of derivatives can be found using the backwards function
d.backwards() # Calculate the derivatives of x*z
print(a.grad) # The gradient of d with respect to a
print(b.grad) # The gradient of b with respect to b