-

import numpy as np
import struct



def read(f, s, offset=0):
    return struct.unpack(f, s[offset:offset+struct.calcsize(f)])


def read_files(imgf, labelf):
    f = open(imgf, 'rb')
    hf = '>iiii'
    _, _, rs, cs = struct.unpack(hf, f.read(struct.calcsize(hf)))
    s = f.read()
    n = len(s) / struct.calcsize('B') / rs / cs
    images = np.ndarray((n, rs*cs, 1), 'B', buffer(s)) / 255.

    f = open(labelf, 'rb')
    hf = '>ii'
    struct.unpack(hf, f.read(struct.calcsize(hf)))
    s = f.read()
    n = len(s) / struct.calcsize('B')
    ls = np.ndarray(n, 'B', buffer(s))
    labels = np.zeros((n, 10, 1))
    for i, l in enumerate(ls):
        labels[i][l][0] = 1.0
    # labels = ls

    return (images, labels)


training_images, training_labels = read_files('train-images.idx3-ubyte', 'train-labels.idx1-ubyte')
test_images, test_labels = read_files('t10k-images.idx3-ubyte', 't10k-labels.idx1-ubyte')


from ch1 import Network

nn = Network([784, 25, 10])

# from ch1 import sigmoid
# a = test_images[0]
# print sigmoid(np.dot(nn.weights[1], sigmoid(np.dot(nn.weights[0], a) + nn.biases[0])) + nn.biases[1])

nn.SGD(
    training_data=zip(training_images, training_labels),
    epochs=30,
    mini_batch_size=10,
    eta=5.0,
    test_data=zip(test_images, test_labels)
)