nn-from-scratch

Neural network implementation from scratch through a Keras-like API in Python.

It is currently a little sketch, thus it just supports:

• Few loss functions: log-loss
• Few activation functions: ReLu & identity
• Few layer types: just fully-connected layers.
• The gradient descent is purely stochastic (batch_size = 1) and implemented using Automatic Differentiation
  • Specifically, forward-propagation is used and implemented using the autograd package.

As per example, the code is applied to solve a classification problem, found at the example.py module. Below, an image of the boundary plot of the built classifier is displayed.

Installation

To install this package's modules into your conda environment conda-env, the .toml file can be leveraged by

(conda-env) $ pip install .

Development (conda) environment

The following needs to be executed in any terminal:

$ conda create -n nn-dev python=3.9 -y
$ conda activate nn-dev
$ conda install numpy sklearn -y
$ conda install -c conda-forge autograd -y
$ conda install -c anaconda sphinx numpydoc \
    sphinx_rtd_theme recommonmark python-graphviz -y
$ pip install --upgrade myst-parser

Documentation

Whenever the modules have been updated, the documentation can be re-generated from the docs folder by typing ():

(nn-dev) nn-from-scratch/docs $ make html