miniPyTorch

This is a toy torch-like framework for educational purposes.
The central idea of this project is to implement autogradient mechanics and show how it works.

Installation

OS X or Linux-like:

chmod +x install.sh
./install.sh

Usage example

To avoid dependency conflict we will keep dev under python virtual environment:

source venv/bin/activate

Let's take a look at how the simplest function can be implemented within our framework and compute it's gradient:

import numpy as np
from mini_torch.tensor import Tensor as T

w1 = T.from_numpy('w1', np.array([-0.91]).astype(float))
w0 = T.from_numpy('w0', np.array([1.5]).astype(float))

x = T.from_numpy('x', np.array([2.]).astype(float), required_grad=False)

y = w0 + w1*x

print(y)

> <class 'mini_torch.tensor.Tensor'>  
> [-0.32]   
> shape: (1,)

y.backward()

print('grad w0', w0.grad)
print('grad w1', w1.grad)

> {'w0': array([1.])}
> {'w1': array([2.])}

Meta

Denistr16 – @github

Distributed under the MIT license. See LICENSE for more information.

https://github.com/denistr16/miniPyTorch

Contributing

1. Fork it (https://github.com/denistr16/miniPyTorch/fork)
2. Create your feature branch (git checkout -b feature/myNewFeature)
3. Commit your changes (git commit -am 'Add some myNewFeature')
4. Push to the branch (git push origin feature/myNewFeature)
5. Create a new Pull Request