backpropagation.py

from random import random
from math import exp


def read_train_file(file="OCR_train.txt"):
    training_data_list = []
    train_file = open(file, "r")
    for line in train_file:
        line = list(line.replace(" ", ""))
        line = [int(x) * 2 - 1 for x in line if x != "\n"]
        training_data_list.extend([line[:]])
    return training_data_list


def active_func_z(z_in):          # activation function using bipolar sigmoid function
    return (1 - exp(-z_in)) / (1 + exp(-z_in))


def set_v(num):
    v = []
    for x in range(num):
        v.extend([[random() for _ in range(64)]])  # initialize weights and biases
        # v.extend([[0] * 64])
    return v


def set_w(num):
    w = []
    for x in range(num):
        w.extend([[random() for _ in range(21)]])  # initialize weights and biases
        # w.extend([[0] * 21])
    return w

cal_eroor = lambda error, total: (error / total) * 100

v = set_v(21)
w = set_w(7)
alpha = 0.1
epoch = 0       # counter of epoch
hidden_unit = 21
training_data = read_train_file()
z_in = []*hidden_unit
y_in = []*7
y = []*7
z = []*hidden_unit
errorOutput = []*7
errorInputIn = []*hidden_unit
errorInput = []*hidden_unit
delta_v = []*hidden_unit
delta_w = []*7


""" TRAINING PHASE OF NN """
# """ 1st PHASE: FEED FORWARD """

# while max(we) > epsilon or ch:                               #check stopping condition
for o in range(9):
    epoch += 1
    ch = False
    for j in training_data:
        x = j[:63]              # set each input unit
        b = j[63]
        expected = j[-7:]
        z_in[0] = 1

        for m in range(hidden_unit):                          # each mid unit
            for V, s in zip(v, x):
                z_in[m] += V * s                     # calculate z_in(j)      j = 1, ..., p
            z[m] = active_func_z(z_in[m])

            for i in range(7):
                for W, z in zip(w, z_in):
                    y_in[i] += W * z
                y[i] = active_func_z(y_in[i])

# """ 2nd PHASE: BACKPROPAGATE ERROR """
            for k in range(7):
                errorOutput[k] = (expected[k] - y[k]) * (1/2 * (1.0 + active_func_z(y_in[k])) * (1.0 - active_func_z(y_in[k])))         # calculate error factor in exit layer
                delta_w[k] = alpha * errorOutput[k] * z[m]          # calculate delta w

            for e in range(hidden_unit):
                for k in range(64):
                    errorInputIn[e] = (expected[k] - y[k]) * (1/2 * (1.0 + active_func_z(y_in[k])) * (1.0 - active_func_z(y_in[k]))) * w[j]         # calculate error factor in mid layer
                errorInput[e] = errorInputIn[e] * active_func_z(z_in[e])
                for X in x:
                    delta_v[e] = alpha * errorInput[e] * X

# """ 3RD PHASE: UPDATE WEIGHTS AND BIASES """

            for n in range(7):          # update weights in exit layer
                w[n] += delta_w[n]

            for n in range(hidden_unit):        # update weights in mid layer
                v[n] += delta_v[n]


print(str(epoch))

"""WEIGHTS AND BIASES SAVING PHASE OF NN"""
weight_file = open("Adeline_weights.txt‬‬", "w")
weight_file.write("Epochs: " + str(epoch) + "th" + "\n" + "\n")
for V in v:
    weight_file.write(str(V) + "\n" + "\n")
weight_file.close()

print("\nThe Neural Network has been trained in " + str(epoch) + "th epochs.")
print("Weights and Biases saved in: ‫Adeline_weights.txt")

"""USE PHASE OF ADELINE NN"""
output = []
_error = 0
_total = 0
results = open("‫‪results_adeline.txt‬‬", "w")
if input("\nDo you want to use your Adeline NN? (y/n)") == 'y':
    test_file = read_train_file("OCR_test.txt")
    for elem in test_file:
        sample = elem[:63]
        target = elem[-7:]
        b=elem[63]

        output.clear()
        _total += 1
        for weight in w:
            result = b
            for w, s in zip(weight, sample):
                result += w * s
            output.append(active_func_z(result))
        if target != output:
            _error += 1
        print("Expected: " + str(target))
        results.write("Expected: " + str(target))
        print("Result:   " + str(output) + "\n------------\n")
        results.write("\nResult:   " + str(output) + "\n------------\n")

print("\n\nPercent of Error in NN: " + str(cal_eroor(_error, _total)))
print("\nNumber of Cells in NN: " + str(len(w)))
results.write("\n\nPercent of Error in NN: " + str(cal_eroor(_error, _total)))
results.write("\nNumber of Cells in NN: " + str(len(w)))
results.close()