A simple pure python implementation of the pytorch API
First, clone this repo.
Install the requirements by running
python -m pip install -r requirements.txt
If you wish to try the interactive demos, also run
python -m pip install -r requirements.extra.txt
Finally, install minitorch by running
python -m pip install -Ue .
I built an autodifferentiation engine with scalars, implementing forward and backward passes for various functions, building a computation graph and utilizing the chain rule, and backpropagation.
While it is not the most efficient implementation, this is enough to train a simple neural network. Here are some examples of classifiers trained based on this simple scalar-based implementation.
To train networks based on this scalar implementation, navigate to the project folder and run
python -m streamlit run app.py -- 1
In this section I am implementing tensors to group operations and improve performance. So far I have implemented tensor indexing and broadcasting and operations, and am currently working on debugging the autodiff system to work for tensors.