https://github.com/ctgk/numgrad
Simple gradient computation library in Python
https://github.com/ctgk/numgrad
Last synced: 3 months ago
JSON representation
Simple gradient computation library in Python
- Host: GitHub
- URL: https://github.com/ctgk/numgrad
- Owner: ctgk
- License: mit
- Created: 2020-08-06T04:14:37.000Z (over 4 years ago)
- Default Branch: main
- Last Pushed: 2024-10-06T02:27:06.000Z (7 months ago)
- Last Synced: 2025-02-01T10:11:31.330Z (3 months ago)
- Language: Python
- Size: 448 KB
- Stars: 3
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# NumGrad
Simple gradient computation library for Python.
# Getting Started
```bash
pip install numgrad
```
Inspired by [tensorflow](https://www.tensorflow.org/), `numgrad` supports [automatic differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation) in tensorflow v2 style using original `numpy` and `scipy` functions.
```python
>>> import numgrad as ng
>>> import numpy as np # Original numpy
>>>
>>> # Pure numpy function
>>> def tanh(x):
... y = np.exp(-2 * x)
... return (1 - y) / (1 + y)
...
>>> x = ng.Variable(1)
>>> with ng.Graph() as g:
... # numgrad patches numpy functions automatically here
... y = tanh(x)
...
>>> g.backward(y, [x])
(0.419974341614026,)
>>> (tanh(1.0001) - tanh(0.9999)) / 0.0002
0.41997434264973155
```
`numgrad` also supports [jax](https://github.com/google/jax) style automatic differentiation.
```python
>>> import numgrad as ng
>>> import numpy as np # Original numpy unlike `jax`
>>>
>>> power_derivatives = [lambda a: np.power(a, 5)]
>>> for _ in range(6):
... power_derivatives.append(ng.grad(power_derivatives[-1]))
...
>>> [f(2) for f in power_derivatives]
[32, 80.0, 160.0, 240.0, 240.0, 120.0, 0.0]
>>> [f(-1) for f in power_derivatives]
[-1, 5.0, -20.0, 60.0, -120.0, 120.0, -0.0]
```
# Contribute
Be sure to run the following command before developing
```bash
$ git clone https://github.com/ctgk/numgrad.git
$ cd numgrad
$ pre-commit install
```