This file pulls down the data set for you. I made a Kaggle user called AustinData. Email me for the password. Data is downloaded to a local folder called data.
This contains the main machine learning that is going on. I use two classes so far -- the FileHandler will grab the data (there is an optional argument for how many lines of data you grab), and the Pixel class is meant to do processing on individual images. For example, the Pixel class has a method self.features() which will return an array of all the features extracted. This is then aggregated in the FileHandler as self.features(), which returns a matrix of the features from each Pixel.
There is also a function normalize which will normalize the features. A good example use of this class is
#!python
data = get_data.get_data_files()
files = analyzer.FileHandler(data['train'])
files.show_bad_classifiers()
get_data_files() will pull down a dictionary with train and test data sets. The FileHandler will train itself, report on how it did, and print a 4 x 4 grid of examples it got wrong. The oob (out of bag) score is a way to not need a cross validation or test set when using a random forest -- it keeps track of which data points went into which decision trees, and excludes those decision trees when predicting for a data point.
The current highest score I've seen is 96.22% correct, which uses each pixel as a feature, as well as the mean of the pixels (Pixel.ink()), which I think should help discriminate between a 1 and an 8, for example.