LeoLionel / simple-mnist

A feed forward Neural Network for MNIST with NumPy

This repository contains the implementation of a simple feed forward neural network for handwritten digit classification on the MNIST dataset.

The code is based on the example in Michael Nielsens (online) book “Neural Networks and Deep Learning”. Additional features are: vectorization over mini batches, Adam optimizer, dropout, different activation and loss functions.

For training download the data (four files) from here and put them in a new subfolder data/, do not unzip them. Now you can train the network with the following code:

import load_mnist
import simple_nn

training_data, test_data = load_mnist.load_data()

net = simple_nn.Network([784,100,10], 'relu')
net.SGD_adam(training_data, test_data,
             epochs = 10,
             mini_batch_size = 10,
             alpha = 0.001, # learning rate
             dropout_rate = 0.1,
             lmbda = 0, # L2 regularization parameter