私は最近それらを研究していたので、python で簡単なニューラル ネットワークを作成しました。バックプロパゲーションを使用しています。活性化関数はシグモイドです。入力と重みはランダムに生成され、理想的な出力は 0 です。私は Python を初めて使用するので、コードは非常に非効率的に書かれていますが、読みやすいです。コードを実行すると、出力は常にゼロになり、その理由がわかりません。また、モジュールは使用しませんでした。
from random import uniform
input_one_input_value = 1
input_two_input_value = 0
bias_value = 1
#use global to globalize function-based variables
#use lists to store data in future
def hidden_One():
global weighted_sum_hidden_one
weighted_sum_hidden_one = (input_one_input_value * weights[0]) + (input_two_input_value * weights[1]) + (bias_value * weights[2])
hidden_one_output = activation(weighted_sum_hidden_one)
return hidden_one_output
def hidden_Two():
global weighted_sum_hidden_two
weighted_sum_hidden_two = (input_one_input_value * weights[3]) + (input_two_input_value * weights[4]) + (bias_value * weights[5])
hidden_two_output = activation(weighted_sum_hidden_two)
return hidden_two_output
def output_One():
weighted_sum = (hidden_One() * weights[6]) + (hidden_Two() * weights[7]) + (bias_value * weights[8])
return activation(weighted_sum)
def activation(x):
sigmoid_value = 1 / (1+(2.71828 ** x))
return sigmoid_value
def calculate_gradient():
E = ideal_output - actual_output
output_delta = -E * activation(weights[6] + weights[7] + weights[8])
h1_delta = activation(weighted_sum_hidden_one) * weights[6] * output_delta
h2_delta = activation(weighted_sum_hidden_two) * weights[7] * output_delta
b2_delta = activation(bias_value) * weights[8] * output_delta
i1_delta = activation(input_one_input_value) * ((weights[0] * h1_delta) + (weights[3] * h2_delta))
i2_delta = activation(input_one_input_value) * ((weights[1] * h1_delta) + (weights[4] * h2_delta))
b1_delta = activation(bias_value) * ((weights[2] * h1_delta) + (weights[5] * h2_delta))
global w1_gradient
global w2_gradient
global w3_gradient
global w4_gradient
global w5_gradient
global w6_gradient
global w7_gradient
global w8_gradient
global w9_gradient
w1_gradient = input_one_input_value * h1_delta
w2_gradient = input_two_input_value * h1_delta
w3_gradient = bias_value * h1_delta
w4_gradient = input_one_input_value * h2_delta
w5_gradient = input_two_input_value * h2_delta
w6_gradient = bias_value * h2_delta
w7_gradient = hidden_One() * output_delta
w8_gradient = hidden_Two() * output_delta
w9_gradient = bias_value * output_delta
def backpropogation():
E = .7 #learning rate
a = .3 #momentum to prevent settling for local minima
global weightchanges_previous
global weight_change
weightchanges_previous = []
weight_change = []
if len(weightchanges_previous) == 0:
weight_change.append((E * w1_gradient))
weight_change.append((E * w2_gradient))
weight_change.append((E * w3_gradient))
weight_change.append((E * w4_gradient))
weight_change.append((E * w5_gradient))
weight_change.append((E * w6_gradient))
weight_change.append((E * w7_gradient))
weight_change.append((E * w8_gradient))
weight_change.append((E * w9_gradient))
weightchanges_previous.append(weight_change[0])
weightchanges_previous.append(weight_change[1])
weightchanges_previous.append(weight_change[2])
weightchanges_previous.append(weight_change[3])
weightchanges_previous.append(weight_change[4])
weightchanges_previous.append(weight_change[5])
weightchanges_previous.append(weight_change[6])
weightchanges_previous.append(weight_change[7])
weightchanges_previous.append(weight_change[8])
elif len(weightchanges_previous) != 0:
weight_change[0] = (E * w1_gradient) + (a * weightchanges_previous[0])
weight_change[1] = (E * w2_gradient) + (a * weightchanges_previous[1])
weight_change[2] = (E * w3_gradient) + (a * weightchanges_previous[2])
weight_change[3] = (E * w4_gradient) + (a * weightchanges_previous[3])
weight_change[4] = (E * w5_gradient) + (a * weightchanges_previous[4])
weight_change[5] = (E * w6_gradient) + (a * weightchanges_previous[5])
weight_change[6] = (E * w7_gradient) + (a * weightchanges_previous[6])
weight_change[7] = (E * w8_gradient) + (a * weightchanges_previous[7])
weight_change[8] = (E * w9_gradient) + (a * weightchanges_previous[8])
while len(weightchanges_previous) > 0 : weightchanges_previous.pop()
weightchanges_previous.append((E * w1_gradient) + (a * weightchanges_previous[0]))
weightchanges_previous.append((E * w2_gradient) + (a * weightchanges_previous[1]))
weightchanges_previous.append((E * w3_gradient) + (a * weightchanges_previous[2]))
weightchanges_previous.append((E * w4_gradient) + (a * weightchanges_previous[3]))
weightchanges_previous.append((E * w5_gradient) + (a * weightchanges_previous[4]))
weightchanges_previous.append((E * w6_gradient) + (a * weightchanges_previous[5]))
weightchanges_previous.append((E * w7_gradient) + (a * weightchanges_previous[6]))
weightchanges_previous.append((E * w8_gradient) + (a * weightchanges_previous[7]))
weightchanges_previous.append((E * w9_gradient) + (a * weightchanges_previous[8]))
def edit_weights():
weights[0] += weight_change[0]
weights[1] += weight_change[1]
weights[2] += weight_change[2]
weights[3] += weight_change[3]
weights[4] += weight_change[4]
weights[5] += weight_change[5]
weights[6] += weight_change[6]
weights[7] += weight_change[7]
weights[8] += weight_change[8]
while len(weight_change) > 0 : weight_change.pop()
weights = []
x = 0
while x <=8:
weights.append(uniform(-10, 10))
x += 1
ideal_output = 0
actual_output = output_One()
print "Output %d" % output_One()
x = 0
while x <= 10:
output_One()
calculate_gradient()
backpropogation()
edit_weights()
print "Output %d" % output_One()
print "----------------------"
actual_output = output_One()
x += 1
print "FINAL WEIGHTS:"
print weights[0]
print weights[1]
print weights[2]
print weights[3]
print weights[4]
print weights[5]
print weights[6]
print weights[7]
print weights[8]