BP算法
基本思想：学习过程由信号的正向传播和误差的反向传播两个过程组成。（这一步体现在propagate()函数）
数学工具：微积分的链式求导法则。（这一步体现在propagate()函数中第34行）
求解最小化成本函数(cost function)：梯度下降法。（这一步体现在optimize()函数）
说明
实现功能：这段代码主要实现的功能是判断一张图片是否有cat，实现的是二分类，有就为1，没有就为0。
训练方法：BP网络，此代码很简单，只有一个神经元，故权值w是一维。（这一步体现在initialiize_with_zeros()中第21行）
难点说明： 第34行：dw = (1./m)*np.dot(X,((A-Y).T)) 此处的dw是指（dL/dw），即损失函数对权值w的导数，此公式是由微积分的链式求导法则推导出来的，吴恩达视频里有，不清楚的请看视频。（2.9 logistic回归中的梯度下降法），这里如果理解了，整个代码也没啥难度了。
注意
1.损失函数和代价函数的区别：
损失函数(Loss function)：指单个训练样本进行预测的结果与实际结果的误差。
代价函数(Cost function)：整个训练集，所有样本误差总和(所有损失函数总和)的平均值。（这一步体现在propagate()函数中的第32行）
1 #!/usr/bin/env python3
2 # -*- coding: utf-8
3
4 import numpy as np
5 import matplotlib.pyplot as plt
6 import h5py
7 import scipy
8 from PIL import Image
9 from scipy import ndimage
10 from lr_utils import load_dataset
11 import pylab
12
13 #sigmoid函数
14 def sigmoid(z):
15 s = 1./(1+np.exp(-z))
16 return(s)
17
18 #初始化权值阈值
19 def initialiize_with_zeros(dim):
20 #这里只有一个神经元，w是一维的
21 w = np.zeros(shape = (dim,1), dtype = np.float32)
22 b = 0
23 #断言函数，判断是否为真
24 assert(w.shape == (dim,1))
25 assert(isinstance(b,float) or isinstance(b,int))
26 return(w,b)
27
28 def propagate(w,b,X,Y):
29 m = X.shape[1]
30 #forward propagation
31 A = sigmoid(np.dot(w.T,X) + b)
32 cost = (-1./m)*np.sum(Y*np.log(A) + (1-Y)*np.log(1-A),axis = 1)#按行相加
33 #backward propagation
34 dw = (1./m)*np.dot(X,((A-Y).T))#dw就是损失函数对w的求导
35 db = (1./m)*np.sum(A-Y, axis=1)#axis=0按列相加，axis=1按行相加
36 assert(dw.shape == w.shape)
37 assert(db.dtype == float)
38 cost = np.squeeze(cost)#squeeze函数的作用是去掉维度为1的维,在这就是将一个一维变成一个数字
39 # [ 6.00006477]
40 # 6.000064773192205
41 assert(cost.shape == ())
42 grads = {"dw": dw,
43 "db": db}
44 return grads, cost
45
46 def optimize(w,b,X,Y,num_iterations,learning_rate,print_cost = False):
47 costs = []
48 for i in range(num_iterations):
49 grads, cost = propagate(w=w, b=b, X=X, Y=Y)
50 dw = grads["dw"]
51 db = grads["db"]
52 w = w - learning_rate*dw
53 b = b - learning_rate*db
54 if i % 100 == 0:
55 costs.append(cost)
56 if print_cost and i % 100 == 0:#这句没懂
57 print ("Cost after iteration %i: %f" %(i, cost))
58 params = {"w": w,
59 "b": b}
60 grads = {"dw": dw,
61 "db": db}
62 return params, grads, costs
63
64 def predict(w, b, X):
65 m = X.shape[1]
66 Y_prediction = np.zeros((1,m))
67 w = w.reshape(X.shape[0], 1)
68 A = sigmoid(np.dot(w.T, X) + b)
69 # [print(x) for x in A]这句没懂，但对代码没啥影响
70 for i in range(A.shape[1]):
71 if A[0, i] >= 0.5:
72 Y_prediction[0, i] = 1
73 else:
74 Y_prediction[0, i] = 0
75 assert(Y_prediction.shape == (1, m))
76
77 return Y_prediction
78
79 def model(X_train, Y_train, X_test, Y_test, num_iterations = 2000, learning_rate = 0.5, print_cost = False):
80 #初始化权值阈值
81 w, b = initialiize_with_zeros(X_train.shape[0])
82 #梯度下降法寻优获取最佳权值阈值
83 parameters, grads, costs = optimize(w, b, X_train, Y_train, num_iterations, learning_rate, print_cost)
84 w = parameters["w"]
85 b = parameters["b"]
86 #测试集进行预测
87 Y_prediction_test = predict(w, b, X_test)
88 Y_prediction_train = predict(w, b, X_train)
89 #输出正确率
90 print("train accuracy: {} %".format(100 - np.mean(np.abs(Y_prediction_train - Y_train)) * 100))
91 print("test accuracy: {} %".format(100 - np.mean(np.abs(Y_prediction_test - Y_test)) * 100))
92 #将所有结果以字典形式保存并返回
93 d = {"costs": costs,
94 "Y_prediction_test": Y_prediction_test,
95 "Y_prediction_train" : Y_prediction_train,
96 "w" : w,
97 "b" : b,
98 "learning_rate" : learning_rate,
99 "num_iterations": num_iterations}
100
101 return d
102
103
104
105
106 ‘‘‘主程序从这里开始‘‘‘
107 #获取训练数据，测试数据
108 train_set_x_orig, train_set_y, test_set_x_orig, test_set_y, classes = load_dataset()
109
110 #reshape()方法来更改数组的形状，train_set_x_orig.shape[0]是行数，-1是代表列数未知，需要numpy自动计算出列数
111 #这里的列数：就是一张图片64*64*3数据变成一行数据的个数
112 train_set_x_flatten = train_set_x_orig.reshape(train_set_x_orig.shape[0], -1).T
113 test_set_x_flatten = test_set_x_orig.reshape(test_set_x_orig.shape[0], -1).T
114
115 #归一，颜色的数值是0~255
116 train_set_x = train_set_x_flatten/255.
117 test_set_x = test_set_x_flatten/255.
118
119 #训练模型
120 d = model(train_set_x, train_set_y, test_set_x, test_set_y, num_iterations = 2000, learning_rate = 0.005, print_cost = True)
121 print(d)
#!/usr/bin/env python3
# -*- coding: utf-8
import numpy as np
import matplotlib.pyplot as plt
import h5py
import scipy
from PIL import Image
from scipy import ndimage
from lr_utils import load_dataset
import pylab
#sigmoid函数
def sigmoid(z):
s = 1./(1+np.exp(-z))
return(s)
#初始化权值阈值
def initialiize_with_zeros(dim):
#这里只有一个神经元，w是一维的
w = np.zeros(shape = (dim,1), dtype = np.float32)
b = 0
#断言函数，判断是否为真
assert(w.shape == (dim,1))
assert(isinstance(b,float) or isinstance(b,int))
return(w,b)
def propagate(w,b,X,Y):
m = X.shape[1]
#forward propagation
A = sigmoid(np.dot(w.T,X) + b)
cost = (-1./m)*np.sum(Y*np.log(A) + (1-Y)*np.log(1-A),axis = 1)#按行相加
#backward propagation
dw = (1./m)*np.dot(X,((A-Y).T))#dw就是损失函数对w的求导
db = (1./m)*np.sum(A-Y, axis=1)#axis=0按列相加，axis=1按行相加
assert(dw.shape == w.shape)
assert(db.dtype == float)
cost = np.squeeze(cost)#????squeeze函数的作用是去掉维度为1的维,在这就是将一个一维变成一个数字
# [ 6.00006477]
# 6.000064773192205
assert(cost.shape == ())
grads = {"dw": dw,
"db": db}
return grads, cost
def optimize(w,b,X,Y,num_iterations,learning_rate,print_cost = False):
costs = []
for i in range(num_iterations):
grads, cost = propagate(w=w, b=b, X=X, Y=Y)
dw = grads["dw"]
db = grads["db"]
w = w - learning_rate*dw
b = b - learning_rate*db
if i % 100 == 0:
costs.append(cost)
if print_cost and i % 100 == 0:#这句没懂
print ("Cost after iteration %i: %f" %(i, cost))
params = {"w": w,
"b": b}
grads = {"dw": dw,
"db": db}
return params, grads, costs
def predict(w, b, X):
m = X.shape[1]
Y_prediction = np.zeros((1,m))
w = w.reshape(X.shape[0], 1)
A = sigmoid(np.dot(w.T, X) + b)
# [print(x) for x in A]
for i in range(A.shape[1]):
# Convert probabilities A[0,i] to actual predictions p[0,i]
### START CODE HERE ### (≈ 4 lines of code)
if A[0, i] >= 0.5:
Y_prediction[0, i] = 1
else:
Y_prediction[0, i] = 0
### END CODE HERE ###
assert(Y_prediction.shape == (1, m))
return Y_prediction
def model(X_train, Y_train, X_test, Y_test, num_iterations = 2000, learning_rate = 0.5, print_cost = False):
#初始化权值阈值
w, b = initialiize_with_zeros(X_train.shape[0])
#梯度下降法寻优获取最佳权值阈值
parameters, grads, costs = optimize(w, b, X_train, Y_train, num_iterations, learning_rate, print_cost)
w = parameters["w"]
b = parameters["b"]
#测试集进行预测
Y_prediction_test = predict(w, b, X_test)
Y_prediction_train = predict(w, b, X_train)
#输出正确率
print("train accuracy: {} %".format(100 - np.mean(np.abs(Y_prediction_train - Y_train)) * 100))
print("test accuracy: {} %".format(100 - np.mean(np.abs(Y_prediction_test - Y_test)) * 100))
#将所有结果以字典形式保存并返回
d = {"costs": costs,
"Y_prediction_test": Y_prediction_test,
"Y_prediction_train" : Y_prediction_train,
"w" : w,
"b" : b,
"learning_rate" : learning_rate,
"num_iterations": num_iterations}
return d
‘‘‘主程序从这里开始‘‘‘
#获取训练数据，测试数据
train_set_x_orig, train_set_y, test_set_x_orig, test_set_y, classes = load_dataset()
#reshape()方法来更改数组的形状，train_set_x_orig.shape[0]是行数，-1是代表列数未知，需要numpy自动计算出列数
#这里的列数：就是一张图片64*64*3数据变成一行数据的个数
train_set_x_flatten = train_set_x_orig.reshape(train_set_x_orig.shape[0], -1).T
test_set_x_flatten = test_set_x_orig.reshape(test_set_x_orig.shape[0], -1).T
#归一，颜色的数值是0~255
train_set_x = train_set_x_flatten/255.
test_set_x = test_set_x_flatten/255.
#训练模型
d = model(train_set_x, train_set_y, test_set_x, test_set_y, num_iterations = 2000, learning_rate = 0.005, print_cost = True)
print(d)