anborde / miniflow

A library that demonstrates training of data using stochastic gradient descent method

Miniflow

A library that demonstrates training of data using stochastic gradient descent method

The minflow library is a part of Udacity's Nanodegree Program and has been prepared while pursuing the same

Table of Contents

• miniflow.py

  This file consists of the set of functions used to perform a basic back propogation in a neural network

• nn.py

  This file consists of a sample neural network

Usage

• Include the miniflow.py in your root project and use the following classess as follows:

  1. Input()

  Use it to declare input nodes of neural network 2. Linear()

  Use to declare a node performing the task of linear activation of form Y = XW+ b 3. Sigmoid()

  Use to declare a node performing sigmodial activation 4. MSE

  Use this node to calculate Mean Square Error

• Further functions are included to perform forward pass, backward pass and gradient descent

• In sample nn file sci-kit learn library is used to resample the dataset to perform a Stochastic Gradient Descent. The miniflow library is designed for Stochastic Gradient Descent and hence the function for gradient descent is named as sgd_update

Contribution Guideline

The miniflow consists only a Linear and a Sigmoidal Activation Nodes. Contributors are expected to provide more depth to the miniflow by including more nodes and functions to perform basic Neural Network based tasks

License

MIT license is included in the repo