fix bug: Loading pickled version of mnist crashes on python 3 #12

[jeffhgs]

Incorporates pickling arguments suggested by maxpumperla https://github.com/maxpumperla/deep_learning_and_the_game_of_go/issues/12

Before submitting a pull request, note that master and the respective chapter branches need to stay in sync with the print version of the book.

• If your PR consists of clarifications, bug fixes, or other vital enhancements, please submit your PR against master.
• If your PR has performance improvements, simplifications, added functionality that would alter the main text of the book etc., please open your PR against the improvements branch.

Should your current PR mix these types of changes, please consider to split it accordingly. Thank you!

[jeffhgs]

Closing:

Under python 2.7 the new unpickling code crashes like so:

  File "/Users/jehenrik/srchome/3pmod/deep_learning_and_the_game_of_go/code/dlgo/nn/load_mnist.py", line 27, in load_data
    data = pickle._Unpickler(f)
AttributeError: 'module' object has no attribute '_Unpickler'

[maxpumperla]

@jeffhgs why not do this:

def load_data():
    with gzip.open('mnist.pkl.gz', 'rb') as f:

        if hasattr(pickle, "_Unpickler"):  # Python 3
            data = pickle._Unpickler(f)
            data.encoding = 'latin1'  # set encoding
            train_data, validation_data, test_data = data.load()
        else:  # Python 2
            train_data, validation_data, test_data = pickle.load(f)