This python project is an implementation of a neural network with a general number of hidden layers. One of the implementations provided in the examples is an handwritten digits classifier.
The file NeuralNetwork.py is a class describing a NeuralNetwork with 1 input layer, n hidden layers, 1 output layer.
It has a backPropagation and a gradientDescent method to train it.
The Main file is using the MNIST database to train a NeuronalNetwork with 30 neurons in the hidden layer, 28*28 neurons in the input layer (1 by pixel) and 10 neurons in the output layer with a stochastic gradient descent algorithm.
It achieve about 95% accuracy on the MNIST testing set
You need to install numpy and python-mnist, this can be done simply with pip. You have to drop the 4 MNIST files in a "Data" folder in the project for the MNIST loader to work (you can get them there : http://yann.lecun.com/exdb/mnist/). The names of the file can change with the versions so make sure they are named only with dashes, no points in the name (if you didn't name them correctly just look at the errors it will tell you what file it was looking for).
You can create a new Neural Network using the constructor, it takes a learning rate and a vector corresponding to the number of neuron in each layer. Example : NeuralNetwork(0.1, [3, 5, 8]) would create a neural network with 3 layers of size 3, 5 and 8 respectively and with learning rate 0.1.
The NeuralNetwork class includes a few helper functions to work on the network such as propagate, back_propagation, gradient_descent, but you can simply use the train function that does all that for you. It takes an exemple input and output, a number of epoch and an epoch size, a test input and output and does all the training and printing of result for you.
Any collaboration is welcome !