from random import randrange
import math

#the first input is a 1 to the bias weight
inputs = [[1,0,0],[1,0,1],[1,1,0],[1,1,1]]
targets = [0,1,1,0]

#the first one in h is the input for the bias weight
h = [1,0.1,0.1]
dh = [1,0,0]
eta = 0.2

def update(perceptron, target, output, x_input, h, dh, eta):
	#find error in output
	dout = (target - output)
	
	#find error in hidden neurons
	for i in range(len(dh)):
		dh[i] = perceptron[1][i]*dout*h[i]*(1-h[i])

	#update output weights
	for i in range(len(perceptron[1])):
		perceptron[1][i] = perceptron[1][i] + eta*dout*h[i]

	#update hidden weights
	for i in range(len(perceptron[0])):
		for j in range(len(perceptron[0][i])):
			perceptron[0][i][j] = perceptron[0][i][j] + eta*x_input[j]*dh[i+1]

def train(perceptron, inputs, targets, h, dh, eta):
	for i in range(5000):
		idx = randrange(len(inputs))
		out = forward(perceptron, inputs[idx], h)
		update(perceptron, targets[idx], out, inputs[idx], h, dh, eta)

	return perceptron

def weightedsum(weights, values):
	return sum(weights[i]*values[i] for i in range(len(weights)))

def forward(perceptron, x_input, h, beta=1):
	for i in range(len(perceptron[0])):
		h[i+1] = 1.0/(1.0+math.exp(-beta*weightedsum(perceptron[0][i], x_input)))
	return weightedsum(perceptron[1],h)

print '\nTRAIN XOR'
#these are random weights. set them to arbitrary small numbers
perceptron = [[[-0.1,-1.1,0.4],[1.1,0.5,-0.6]], [0.7,0.3,-0.5]]
perceptron = train(perceptron, inputs, targets, h, dh, eta)

print "[0,0] ", forward(perceptron, inputs[0], h)
print "[0,1] ", forward(perceptron, inputs[1], h)
print "[1,0] ", forward(perceptron, inputs[2], h)
print "[1,1] ", forward(perceptron, inputs[3], h)