BP網絡——前饋神經網絡
(Back Propgation Networks)
本文將實現一個最簡單的三層神經網絡
其中,損失函數是線性函數,激勵函數是sigmoid函數。
代碼的實現中,採用的是隨機梯度下降法。
計算導數的方法可以參考圖片,有興趣的小夥伴可以參考,自行推導:
隨機梯度下降法的原理
代碼如下所示:
#!/user/bin/env python3
# -*- coding: utf-8 -*-
import random
import numpy as np
#定義神經網絡結構
class Network(object):
def __init__(self, sizes):
# 網絡層數
self.num_layers = len(sizes)
# 每層神經元的個數
self.sizes = sizes
# 初始化每層的偏置
self.biases = [np.random.randn(y, 1) for y in sizes[1:]]
# 初始化每層的權重
self.weights = [np.random.randn(y, x) for x, y in zip(sizes[:-1], sizes[1:])]
def feedforward(self, a):
for b, w in zip(self.biases, self.weights):
a = sigmoid(np.dot(w, a) + b)
return a
# 梯度下降
def SGD(self, training_data, epochs, mini_batch_size, eta, test_data=None):
if test_data:
n_test = len(test_data)
# 訓練數據總個數
n = len(training_data)
# 開始訓練,循環每一個epochs
for j in range(epochs): # 在python2.7中爲xrange
# 洗牌 打亂訓練數據
random.shuffle(training_data)
# mini_batch
mini_batches = [training_data[k:k+mini_batch_size] for k in range(0, n, mini_batch_size)]
# 訓練mini_batch
for mini_batch in mini_batches:
self.update_mini_batch(mini_batch, eta)
if test_data:
print("Epoch {0}: {1} / {2}".format(j, self.evaluate(test_data), n_test))
print("Epoch {0} complete".format(j))
def update_mini_batch(self, mini_batch, eta):
# 保存每層偏導
nabla_b = [np.zeros(b.shape) for b in self.biases]
nabla_w = [np.zeros(w.shape) for w in self.weights]
# 訓練一個mini_batch
for x, y in mini_batch:
delta_nabla_b, delta_nabla_w = self.update(x, y)
# 保存一次訓練網絡中每層的偏導
nabla_b = [nb + dnb for nb, dnb in zip(nabla_b, delta_nabla_b)]
nabla_w = [nw + dnw for nw, dnw in zip(nabla_w, delta_nabla_w)]
# 更新權重和偏置 Wn+1 = Wn - eta * nw
self.weights = [w - (eta / len(mini_batch)) * nw for w, nw in zip(self.weights, nabla_w)]
self.biases = [b - (eta / len(mini_batch)) * nb for b, nb in zip(self.biases, nabla_b)]
# 前向傳播
def update(self, x, y):
# 保存每層偏導
nabla_b = [np.zeros(b.shape) for b in self.biases]
nabla_w = [np.zeros(w.shape) for w in self.weights]
activation = x # 保存的輸入(訓練數據)
# 保存每一層的激勵值a=sigmoid(z)
activations = [x]
# 保存每一層的z=wx+b
zs = []
# 前向傳播
for b, w in zip(self.biases, self.weights):
# 計算每層的z
z = np.dot(w, activation) + b
# 保存每層的z
zs.append(z)
# 計算每層的a
activation = sigmoid(z)
# 保存每一層的a
activations.append(activation)
# 反向更新
# 計算最後一層的誤差
delta = self.cost_derivative(activations[-1], y) * sigmoid_prime(zs[-1])
# 最後一層權重和偏置的導數
nabla_b[-1] = delta
nabla_w[-1] = np.dot(delta, activations[-2].T)
# 倒數第二層一直到第一層 權重和偏置的導數
for l in range(2, self.num_layers):
z = zs[-l]
sp = sigmoid_prime(z)
# 當前層的誤差
delta = np.dot(self.weights[-l+1].T, delta) * sp
# 當前層的偏置和權重的導數
nabla_b[-l] = delta
nabla_w[-l] = np.dot(delta, activations[-l-1].T)
return (nabla_b, nabla_w)
def evaluate(self, test_data):
test_results = [(np.argmax(self.feedforward(x)), y) for (x, y) in test_data]
return sum(int(x == y) for x, y in test_results)
def cost_derivative(self, output_activation, y):
return (output_activation - y)
# sigmoid激勵函數
def sigmoid(z):
return 1.0 / (1.0 + np.exp(-z))
def sigmoid_prime(z):
return sigmoid(z) * (1-sigmoid(z))
if __name__ == "__main__":
import mnist_loader
training_data, validation_data, test_data = mnist_loader.load_data_wrapper()
# 28*28=784個像素, 可以定義30個神經元, 共有10種分類
net = Network([784, 30, 10])
net.SGD(training_data, 30, 10, 0.5, test_data=test_data)
由於進行預測的時候用的是mnist數據集
所以需要下載mnist.pkl.gz
mnist.pkl.gz下載地址爲:mnist.pkl.gz
小夥伴千萬不要下錯哦,之前我下的就是個數據不夠的,結果害得我查了好久爲什麼代碼會出錯,哭唧唧。。。
注:mnist_loader.py是一個python文件(如下)
#!/user/bin/env python3
# -*- coding: utf-8 -*-
"""
mnist_loader
~~~~~~~~~~~~
A library to load the MNIST image data. For details of the data
structures that are returned, see the doc strings for ``load_data``
and ``load_data_wrapper``. In practice, ``load_data_wrapper`` is the
function usually called by our neural network code.
"""
#### Libraries
# Standard library
import pickle
import gzip
import numpy as np
def load_data():
f = gzip.open('data/mnist.pkl.gz', 'rb')
training_data, validation_data, test_data = pickle.load(f, encoding='bytes')
f.close()
return (training_data, validation_data, test_data)
def load_data_wrapper():
tr_d, va_d, te_d = load_data()
training_inputs = [np.reshape(x, (784, 1)) for x in tr_d[0]]
training_results = [vectorized_result(y) for y in tr_d[1]]
training_data = list(zip(training_inputs, training_results))
validation_inputs = [np.reshape(x, (784, 1)) for x in va_d[0]]
validation_data = list(zip(validation_inputs, va_d[1]))
test_inputs = [np.reshape(x, (784, 1)) for x in te_d[0]]
test_data = list(zip(test_inputs, te_d[1]))
return (training_data, validation_data, test_data)
def vectorized_result(j):
e = np.zeros((10, 1))
e[j] = 1.0
return e
大家一定要注意python版本的問題,我用的是python3,和python2還是有差別的,比如以下幾點:
①python3中沒有xrange,python3中的range實際上等同於python2中的xrange。
②好像pickle的使用也有所不同,在python2中是cpickle,但python3中是pickle;
很多人出現了一個問題,就是python3使用pickle.load()方法時,需要加上pickle.load(f, encoding='bytes')
encoding編碼方式才能解決。
③python2中的zip方法和python3不同,python3的zip()不是列表,所以在上述代碼的load_data_wrapper()方法中,zip出現的地方需要加上list()方法。
注:
代碼仍然有很多可以進行提升的部分,見下一篇博客:反向傳播(BP)網絡的mnist實例(提升篇)