vanilla_nn.py

import random
import itertools


def E(x, y):
    return (x - y) ** 2


def forward_pass(w, x):
    x[2] = w[0] * x[0] + w[2] * x[1]
    x[3] = w[1] * x[0] + w[3] * x[1]
    x[4] = w[4] * x[2] + w[5] * x[3]
    return x


def backward_pass(w, x, y, l=0.05):
    dEdx_4 = -2 * (y - x[4])
    dx_4dw_4 = x[2]
    dx_4dw_5 = x[3]
    dx_4dx_2 = w[4]
    dx_4dx_3 = w[5]
    dx_2dw_0 = x[0]
    dx_2dw_2 = x[1]
    dx_3dw_1 = x[0]
    dx_3dw_3 = x[1]

    dEdw = [0] * 6
    dEdw[5] = dEdx_4 * dx_4dw_5
    dEdw[4] = dEdx_4 * dx_4dw_4
    dEdw[3] = dEdx_4 * dx_4dx_3 * dx_3dw_3
    dEdw[2] = dEdx_4 * dx_4dx_2 * dx_2dw_2
    dEdw[1] = dEdx_4 * dx_4dx_3 * dx_3dw_1
    dEdw[0] = dEdx_4 * dx_4dx_2 * dx_2dw_0

    dw = [-l * w_i for w_i in dEdw]
    w = [w_i + dw_i for w_i, dw_i in zip(w, dw)]
    return w


xs = [(1, 2), (1.5, 3), (2, 3), (2, 1), (3, 2), (3, 1)]
ys = [1, 1, 1, -1, -1, -1]

w = [round(random.random(), 2) for _ in range(6)]
print('initial weights:', w)

for x_i, y_i in zip(xs, ys):
    x = [x_i[0], x_i[1], 0, 0, 0]
    predicted = forward_pass(w, x)[4]
    print('predicted:', predicted, 'actual:', y_i)

print('training...')

epoch = 0
for x_i, y_i in itertools.cycle(zip(xs, ys)):
    x = [x_i[0], x_i[1], 0, 0, 0]
    y = y_i

    # print('input vector:', x)
    x = forward_pass(w, x)
    # print('x after forward pass:', x)
    error = E(x[4], y)
    # print('error', error)
    w = backward_pass(w, x, y)
    # print('w after backwards pass', w)

    epoch += 1
    if epoch > 100:
        break

for x_i, y_i in zip(xs, ys):
    x = [x_i[0], x_i[1], 0, 0, 0]
    predicted = forward_pass(w, x)[4]
    print('predicted:', predicted, 'actual:', y_i)