junlulocky / classiccvproblems

Classic computer vision problems (MNIST, CIFAR10, SVHN)

Hand Written Recognition

This repository contains my test of classic handwritten recognition problem on MNIST.

The dataset contains the MNIST dataset, which is from http://www.iro.umontreal.ca/~lisa/deep/data/mnist/mnist.pkl.gz.

The start code is from theano tutorial, but my implementation goes beyond it's scope.

Code Structure

/MNIST/theano

|-layers
    |-hiddenLayer.py    # hidden layer
    |-convPoolLayer.py  # convolutional + max pooling layer
    |-logisticRegLayer.py  # usually used for the output

|-update
    |-activationFunctions.py  # non-linear functions used in NN
    |-updateMethods.py        # implementation of update methods

|-loader.py   # load data
|-mainLogisticRegression.py  # classify by logistic regression
|-mainMLP.py  # classify by MLP
|-mainCNN.py  # classify by CNN - classic LeNet architecture  

/MNIST

The code relies on an unpublished library, so it cannot run. It is just a NN or CNN structure.

|-train_mnist_mlp.py   # use MLP to solve MNIST dataset
|-train_mnist_lenet.py # use lenet structure to solve MNIST dataset

/CIFAR10

links for CIFAR10:

• From lasagne
• From binary net

train_cifar10_fpwider.py  # Floating Point structure, with flipping, less than 15% error rate

The code uses torch,

|-cifar10_lenet.lua   # use LeNet structure to solve CIFAR10 dataset

/SVHN

The code relies on an unpublished library, so it cannot run. It is just a NN or CNN structure.

|-train_svhn_fpnet.py # Floating point structure, less than 3.3% error rate.