经典篇
Step1:导入需要的模块
其中，h5py模块用于读取.h5数据（本次的训练数据和测试数据均为.h5文件），lr_utils包含数据加载函数，为官方提供的模块。
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 math
Step2:导入数据-分析数据-标准化数据
可以看到训练数据共209个样本，验证数据共50个样本，样本图片为RGB数据，维度为64*64*3。我们首先将输入数据reshape为（64*64*3，none），none对于训练集为209，对于验证集为50。随后将输入数据标准化，因为图片的像素值在0-255之间，我们将其转换为0-1之间的小数进而方便我们计算。
# Input data
train_set_x_orig, train_set_y, test_set_x_orig, test_set_y, classes = load_dataset()
train_num = train_set_x_orig.shape[0]
test_num = test_set_x_orig.shape[0]
pix_num = train_set_x_orig.shape[1]
chanel_num = train_set_x_orig.shape[3]
print('\n' + '---------------Input information---------------' + '\n')
print('train_set_x_orig.shape = ' + str(train_set_x_orig.shape))
print('train_set_y.shape = ' + str(train_set_y.shape))
print('test_set_x_orig.shape = ' + str(test_set_x_orig.shape))
print('test_set_y.shape = ' + str(test_set_y.shape))
print('train_num = ' + str(train_num))
print('test_num = ' + str(test_num))
print('pix_num = ' + str(pix_num))
print('chanel_num = ' + str(chanel_num))
# Reshape data
train_set_x_flat = train_set_x_orig.reshape(train_num,-1).T
test_set_x_flat = test_set_x_orig.reshape(test_num,-1).T
print('\n' + '---------------After reshaping---------------' + '\n')
print('train_set_x_flat = ' + str(train_set_x_flat.shape))
print('test_set_x_flat = ' + str(train_set_x_flat.shape))
# Standarize data
train_set_x = train_set_x_flat/255.0
test_set_x = test_set_x_flat/255.0
print('\n' + '---------------After Standaring---------------' + '\n')
print('Check for traindata = ' + str(train_set_x[0:5,0]))
print('Check for testdata = ' + str(test_set_x[0:5,0]))
Step3:定义激活函数及其导数
主要包括sigmoid函数，tanh函数，relu函数及其对应的导数，实际操作根据具体需要选择合适的激活函数（详见 我的另一篇博客）。
def sigmoid(Z):
s = 1.0/(1.0 + np.exp(-Z))
return s
def sigmoid_prime(Z):
s = sigmoid(Z)*(1-sigmoid(Z))
return s
def tanh(x):
s = (np.exp(x)-np.exp(-x))/((np.exp(x)+np.exp(-x)))
return s
def tanh_prime(x):
s = 1 - np.power(tanh(x), 2)
return s
def relu(Z):
s = np.maximum(0,Z)
return s
def relu_prime(Z):
s = Z
s[Z <= 0] = 0
s[Z>0]=1
return s
Step4:初始化函数-用于初始化权重
def initial_weights(layer):
# layer :[nx, n1, n2, ... nL]
np.random.seed(2)
L = len(layer)
parameters = {}
for l in range(1,L):
parameters["W"+str(l)] = np.random.randn(layer[l], layer[l-1])*np.sqrt(2.0/layer[l-1])
parameters["b"+str(l)] = np.zeros((layer[l],1))
return parameters
Step5:训练函数
def train(X, Y, layer, learning_rate = 0.0075, itr_num = 3000,isprint=True):
'''''
Input:
X: training dataset with shape(nx,m)
Y: training dataset with shape(1,m)
learning_rate: default with 0.0075
itr_num: max number of iteration
isprint: print cost(True) or not(False)
Output:
parameters : dict data{"W1","b1",...,"WL","bL"},include final weights
'''''
m = Y.shape[1] # number of training set
Fcahe = {} # Save forward propagation variable(Z,A)
Bcahe = {} # Save backward propagation variable(dZ,dW,db)
costs = [] # Save cost series to plot
Fcahe["A"+str(0)] = X # Set X as A0
parameters = initial_weights(layer) # Initial weights
L = len(parameters) // 2 # Layers of neural netwoek
for itr in range(itr_num):
# Forward propagation
for l in range(1,L):
Fcahe["Z"+str(l)] = np.dot(parameters["W"+str(l)], Fcahe["A"+str(l-1)]) + parameters["b"+str(l)]
Fcahe["A"+str(l)] = relu(Fcahe["Z"+str(l)])
Fcahe["Z"+str(L)] = np.dot(parameters["W"+str(L)], Fcahe["A"+str(L-1)]) + parameters["b"+str(L)]
Fcahe["A"+str(L)] = sigmoid(Fcahe["Z"+str(L)])
cost = -(1.0/m)*np.sum(Y*np.log( Fcahe["A"+str(L)])+(1-Y)*np.log(1 - Fcahe["A"+str(L)]))
cost = np.squeeze(cost)
# Backward propagation
Bcahe["dZ"+str(L)]=(-(np.divide(Y,Fcahe["A"+str(L)])-np.divide(1-Y,1-Fcahe["A"+str(L)])))*sigmoid_prime(Fcahe["Z"+str(L)])
Bcahe["dW"+str(L)]=(1.0/m)*np.dot(Bcahe["dZ"+str(L)],Fcahe["A"+str(L-1)].T)
Bcahe["db"+str(L)]=(1.0/m)*np.sum(Bcahe["dZ"+str(L)],axis=1,keepdims=True)
for l in reversed(range(L-1)):
Bcahe["dZ"+str(l+1)]= np.dot(parameters["W"+str(l+2)].T, Bcahe["dZ"+str(l+2)])*relu_prime(Fcahe["Z"+str(l+1)])
Bcahe["dW"+str(l+1)]=(1.0/m)*np.dot(Bcahe["dZ"+str(l+1)],Fcahe["A"+str(l)].T)
Bcahe["db"+str(l+1)]=(1.0/m)*np.sum(Bcahe["dZ"+str(l+1)],axis=1,keepdims=True)
# Update weights
for l in range(L):
parameters["W"+str(l+1)] = parameters["W"+str(l+1)] -learning_rate*Bcahe["dW"+str(l+1)]
parameters["b"+str(l+1)] = parameters["b"+str(l+1)] - learning_rate*Bcahe["db"+str(l+1)]
if isprint and itr%200 == 0:
print('cost after ' + str(itr) + ' iteration is :' + str(cost))
costs.append(cost)
# Plot the cost curve vary with itr_num
plt.plot(np.squeeze(costs))
plt.xlabel("itr_num(per hundred)")
plt.ylabel("cost")
plt.title("Learning rate =" + str(learning_rate))
plt.show()
return parameters
Step6:测试函数
def predict(X,Y,parameters):
'''''
Input:
X: test dataset with shape(nx,m)
Y: test dataset with shape(1,m)
parameters : dict data{"W1","b1",...,"WL","bL"}
Output:
pred: predicted labels of test dataset
'''''
num = X.shape[1]
pred = np.zeros((1,num))
L = len(parameters) // 2
A0 = X
for l in range(1,L):
Z = np.dot(parameters["W"+str(l)],A0) +parameters["b"+str(l)]
A = relu(Z)
A0 = A
Z = np.dot(parameters["W"+str(L)],A0) +parameters["b"+str(L)]
A = sigmoid(Z)
for i in range(A.shape[1]):
if A[0,i]<=0.5:
pred[0,i]=0
else:
pred[0,i]=1
print("accuracy:{}%".format(100 - np.mean(np.abs(pred - Y)) * 100))
return pred
Step6:代码测试
测试1：Logistic Regression 测试，包含输入层和输出层，输入层个数为12288，输出层节点个数为1，学习率为0.01，共迭代3000次。三次测试的随机种子设置为np.random.seed(3)，最终验证集准确率为68%。
parameters = train(train_set_x, train_set_y, [12288,1], learning_rate= 0.01, itr_num=3000)
test_pred = predict(test_set_x,test_set_y,parameters)
测试2：单隐层神经网络测试，包含输入层、隐层以及输出层，输入层个数为12288，隐层节点个数为5，输出层节点个数为1，学习率为0.01，共迭代3000次，最终验证集准确率为70%。
parameters = train(train_set_x, train_set_y, [12288,5,1], learning_rate= 0.01, itr_num=3000)
test_pred = predict(test_set_x,test_set_y,parameters)
测试3：三层神经网络测试，包含输入层、隐层1、隐层2以及输出层，输入层个数为12288，隐层1节点个数为20，隐层2节点个数为5，输出层节点个数为1，学习率为0.01，共迭代3000次，最终验证集准确率为78%。
parameters = train(train_set_x, train_set_y, [12288,20,5,1], learning_rate= 0.01, itr_num=3000)
test_pred = predict(test_set_x,test_set_y,parameters)
|
转播
