mnist_detection.md

MNIST detection

These sections includes a breif description of the basic usage for the MNIST detection implementation.

Data sets

The implementation uses a synthetic dataset created from the original MNIST images. It is created by running:

$ th create_datasets_72.lua
$ th extract crops.lua

This will download the original MNIST dataset, create files mnist_72.t7 and mnist_72_test.t7, extract crops from those and store them as mnist_train.t7 and mnist_test.t7.

Training

The script main_cuda.lua contains all parameters used to declare the training and a high-level training function train(). Make sure to set up the absolute path to the repository:

path = '/......./masterThesis/src/mnist_detection/'

Testing

The network accuracy is tested with the evaluateError() function, taking arguments 'validation' or 'test'. Training cost can be plotted with function plotCost().

The file plotfunc_cuda.lua contains various functions used to test the network performance qualitatively.