import numpy as np
import time
import sys

class NeuralNetwork:
    def __init__ (self, layers, eps = 0.3):
        self.w = []
        self.eta = 0.3
        self.learningStepCount = 500
        for i in range(1, len(layers)):
            self.w.append(np.matrix(eps * (2 * np.random.random((layers[i], layers[i - 1])) -  1)))

    def sigmoid(self, x):
        return 1.0 / (1 + np.exp(-x))

    def getOutput(self, x):
        for layer in self.w:
            x = self.sigmoid(layer * x)
        ind = x.argmax()
        x = np.zeros(len(x))
        x[ind] = 1
        return x

    def gradient(self, x):
        return np.multiply(x, 1 - x)
    #@profile
    def train(self, testX, testY):
        for step in range(self.learningStepCount):
            for xInput, y in zip(testX, testY):
                x = []
                for layer in self.w:
                    x.append(xInput)
                    xInput = self.sigmoid(layer * xInput)

                output = xInput
                delta = np.multiply(self.gradient(output), (y - output))

                for i in reversed(range(len(self.w))):
                    self.w[i] += self.eta * delta * x[i].T
                    delta = np.multiply(self.gradient(x[i]), self.w[i].T * delta)
            print(step)
#
@profile
def main():
    x, y = [], []
    tests = []
    f = open('train.csv', 'r')
    for line in f:
        tests.append(np.array(list(map(int, line.split(',')))))
    f.close()

    np.random.shuffle(tests)
    for test in tests:
        x.append(np.matrix(test[1:]).T)
        tmp = [0] * 10
        tmp[int(test[0]) - 1] = 1
        y.append(np.matrix(tmp).T)

    netw = NeuralNetwork([784, 2000, 10])
    tests = tests[:len(tests) // 10]
    testCount = len(tests)
    print(testCount)
    tests.clear()

    netw.train(x[:testCount // 2], y[:testCount // 2])
    good = 0
    for i in range(testCount // 2, testCount):
        if (netw.getOutput(x[i]) == y[i]).all():
            good += 1

    print('Spent', time.time() - timer, 'seconds')
    print('Accuracy:', good * 1.0 / testCount * 2)

if __name__ == '__main__':
    main()