iamtrask_recurrentNN_tutorial

#link to tutorial: https://iamtrask.github.io/2015/11/15/anyone-can-code-lstm/
import copy, numpy as np
np.random.seed(0)
# compute sigmoid nonlinearity
def sigmoid(x):
    output = 1/(1+np.exp(-x))
    return output

# convert output of sigmoid function to its derivative
def sigmoid_output_to_derivative(output):
    return output*(1-output)

def leaky_relu(x):
    if x >= 0:
        return x
    else:
        return x / 100.0

def relu_output_to_derivative(output):
    if output >= 0:
        return 1
    else:
        return 0.01

def int2trinary(intgr):
    remainder = intgr
    trinary = []
    for i in range(trinary_dim,-1,-1):
        trinary.append(remainder/pow(3,i))
        remainder = remainder%pow(3,i)
    return trinary



# training dataset generation
trinary_dim = 5
largest_number = pow(3,trinary_dim)

# input variables
alpha = 0.1
input_dim = 2
hidden_dim = 10
output_dim = 1


# initialize neural network weights
synapse_0 = 2*np.random.random((input_dim,hidden_dim)) - 1
synapse_1 = 2*np.random.random((hidden_dim,output_dim)) - 1
synapse_h = 2*np.random.random((hidden_dim,hidden_dim)) - 1

synapse_0_update = np.zeros_like(synapse_0)
synapse_1_update = np.zeros_like(synapse_1)
synapse_h_update = np.zeros_like(synapse_h)

# training logic
for j in range(70000):

    # generate a simple addition problem (a + b = c)
    a_int = np.random.randint(largest_number/2) # int version
    a = int2trinary(a_int) # binary encoding

    b_int = np.random.randint(largest_number/2) # int version
    b = int2trinary(b_int) # binary encoding

    # true answer
    c_int = a_int + b_int
    c = int2trinary(c_int)

    # where we'll store our best guess (binary encoded)
    d = np.zeros_like(c)

    overallError = 0

    layer_2_deltas = list()
    layer_1_values = list()
    layer_1_values.append(np.zeros(hidden_dim))

    # moving along the positions in the binary encoding
    for position in range(trinary_dim):

        # generate input and output
        X = np.array([[a[trinary_dim - position - 1],b[trinary_dim - position - 1]]])
        y = np.array([[c[trinary_dim - position - 1]]]).T

        # hidden layer (input ~+ prev_hidden)
        layer_1 = sigmoid(np.dot(X,synapse_0) + np.dot(layer_1_values[-1],synapse_h))

        # output layer (new binary representation)
        layer_2 = leaky_relu(np.dot(layer_1,synapse_1))

        # did we miss?... if so, by how much?
        layer_2_error = y - layer_2
        layer_2_deltas.append((layer_2_error)*relu_output_to_derivative(layer_2))
        overallError += np.abs(layer_2_error[0])

        # decode estimate so we can print it out
        d[trinary_dim - position - 1] = np.round(layer_2[0][0])

        # store hidden layer so we can use it in the next timestep
        layer_1_values.append(copy.deepcopy(layer_1))

    future_layer_1_delta = np.zeros(hidden_dim)

    for position in range(trinary_dim):

        X = np.array([[a[position],b[position]]])
        layer_1 = layer_1_values[-position-1]
        prev_layer_1 = layer_1_values[-position-2]

        # error at output layer
        layer_2_delta = layer_2_deltas[-position-1]
        # error at hidden layer
        layer_1_delta = (future_layer_1_delta.dot(synapse_h.T) + layer_2_delta.dot(synapse_1.T)) * sigmoid_output_to_derivative(layer_1)

        # let's update all our weights so we can try again
        synapse_1_update += np.atleast_2d(layer_1).T.dot(layer_2_delta)
        synapse_h_update += np.atleast_2d(prev_layer_1).T.dot(layer_1_delta)
        synapse_0_update += X.T.dot(layer_1_delta)

        future_layer_1_delta = layer_1_delta


    synapse_0 += synapse_0_update * alpha
    synapse_1 += synapse_1_update * alpha
    synapse_h += synapse_h_update * alpha

    synapse_0_update *= 0
    synapse_1_update *= 0
    synapse_h_update *= 0

    # print out progress
    if(j % 10000 == 0):
        print "Error:" + str(overallError)
        print "Pred:" + str(d)
        print "True:" + str(c)
        out = 0
        for index,x in enumerate(reversed(d)):
            out += x*pow(3,index)
        print str(a_int) + " + " + str(b_int) + " = " + str(out)
        print "------------"