tensorflow入门实例

一、回归算法

import tensorflow as tf
import numpy as np

#-------------------1. 数据集，变量，占位符------------------------#

# 样本，输入列表，正太分布(Normal Destribution)，均值为1, 均方误差为0.1, 数据量为100个
x_vals = np.random.normal(1, 0.1, 100)
# 样本, 输出列表， 100个值为10.0的列表
y_vals = np.repeat(10.0, 100)

#占位符
x_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
y_target = tf.placeholder(shape=[None, 1], dtype= tf.float32)

#模型变量
A = tf.Variable(tf.random_normal(shape=[1, 1]))

#批量大小
batch_size = 25

#训练数据集的index，从总样本的index，即0～99,选取80个值
train_indices = np.random.choice(len(x_vals), round(len(x_vals) *0.8), replace = False)
#测试数据集的index，扣除上面的train_indices，剩下的20个值
test_indices = np.array(list(set(range(len(x_vals))) - set(train_indices)))

#训练数据集 & 测试数据集
x_vals_train = x_vals[train_indices]
x_vals_test = x_vals[test_indices]
y_vals_train = y_vals[train_indices]
y_vals_test = y_vals[test_indices]

#-----------------2. 模型，损失函数，优化器算法--------------------------#

# 我们定义的模型，是一个线型函数，即 y = w * x， 也就是my_output = A * x_data
# x_data将用样本x_vals。我们的目标是，算出A的值。
# 其实已经能猜出，y都是10.0的话，x均值为1, 那么A应该是10。哈哈
my_output = tf.multiply(x_data, A)

# 损失函数， 用的是模型算的值，减去实际值， 的平方。y_target就是上面的y_vals。
loss = tf.reduce_mean(tf.square(my_output - y_target))

#初始化变量
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

# 梯度下降算法， 学习率0.02, 可以认为每次迭代修改A，修改一次0.02。比如A初始化为20, 发现不好，于是猜测下一个A为20-0.02
my_opt = tf.train.GradientDescentOptimizer(learning_rate=0.02)
train_step = my_opt.minimize(loss)#目标，使得损失函数达到最小值


#-----------------3. 迭代训练--------------------------#


for i in range(100):#0到100,不包括100
    # 随机拿25个index
    rand_index = np.random.choice(len(x_vals_train), size = batch_size)
    # 从训练集拿出25个样本，转置一下，因为x_data的shape是[None, 1]
    #注意是[x_vals_train[rand_index]]，转为二维的1x20的数组，才能通过transpose转置为20x1的数组，不能写成x_vals_train[rand_index]
    rand_x = np.transpose([x_vals_train[rand_index]])
    rand_y = np.transpose([y_vals_train[rand_index]])
    #损失函数引用的placeholder(直接或间接用的都算), x_data使用样本rand_x， y_target用样本rand_y
    sess.run(train_step, feed_dict={x_data: rand_x, y_target: rand_y})
    #打印
    if i%25==0:
        print('step: ' + str(i) + ' A = ' + str(sess.run(A)))
        print('loss: ' + str(sess.run(loss, feed_dict={x_data: rand_x, y_target: rand_y})))

#-----------------4. 评估模型--------------------------#
#以上这种评估，测试集跟训练集是完全分开的。没有用A去评测测试集，只看两种集的均方误差是不是差不多
mse_test = sess.run(loss, feed_dict={x_data: np.transpose([x_vals_test]), y_target: np.transpose([y_vals_test])})
mse_train = sess.run(loss, feed_dict={x_data: np.transpose([x_vals_train]), y_target: np.transpose([y_vals_train])})
print('MSE on test: ' + str(np.round(mse_test, 2)))
print('MSE on train: ' + str(np.round(mse_train, 2)))

 

二、分类算法

#import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
#sklearn是机器学习套件，有很多数据集
#安装：pip install -U scikit-learn
#      sklearn依赖python>=2.7, numpy(python擅长数组处理的数学库), scipy(python算法库和数据工具包)
from sklearn import datasets

#-------------------1. 数据集，变量，占位符------------------------#
iris = datasets.load_iris()
print('sample feature: feature_names: ' + str(iris.feature_names) + " data length: " + str(len(iris.data)))
print('sample target: target_names: ' + str(iris.target_names) + " target length: " + str(len(iris.target)))
#样本数据，一个150x4的二维列表
#print(iris.data)
#样本标签，一个长度为150的一维列表
#print(iris.target)

#抽取的样本标签, 只要第一种，是第一种，则为1，否则为0
temp = []
for x in iris.target:
    temp.append(1 if x== 0 else 0)
iris_target = np.array(temp)#列表转数组，以上几行，也可以写成：iris_target = np.array([1 if x== 0 else 0 for x in iris.target])
print('iris_target: ')
print(iris_target)

#抽取的样本输入，只用两个参数，也就是花瓣长度和宽度
iris_2d = np.array([[x[2], x[3]] for x in iris.data])
print('iris_2d: ')
print(iris_2d)

#将样本分为训练集和测试集
#训练数据集的index，从总样本的index，即0～150,选取120个值
train_indices = np.random.choice(len(iris_2d), round(len(iris_2d) *0.8), replace = False)
#测试数据集的index，扣除上面的train_indices，剩下的30个值
test_indices = np.array(list(set(range(len(iris_2d))) - set(train_indices)))

#训练数据集 & 测试数据集
x_vals_train = iris_2d[train_indices]
x_vals_test = iris_2d[test_indices]
y_vals_train = iris_target[train_indices]
y_vals_test = iris_target[test_indices]

#批量训练大小为20
batch_size = 20
x1_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
x2_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
y_target = tf.placeholder(shape=[None, 1], dtype=tf.float32)

A = tf.Variable(tf.random_normal(shape=[1,1]))
b = tf.Variable(tf.random_normal(shape=[1,1]))

#-----------------2. 模型，损失函数，优化器算法--------------------------#
#定义模型my_output = x1 - (A*x2 + b)
my_mult = tf.matmul(x2_data, A)
my_add = tf.add(my_mult, b)
my_output = tf.subtract(x1_data, my_add)

#损失函数
xentropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=my_output, logits=y_target)
my_opt = tf.train.GradientDescentOptimizer(0.05)
train_step = my_opt.minimize(xentropy)

#初始化变量
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

#-----------------3. 迭代训练--------------------------#

#开始迭代，更新模型，也就是计算出A和b
for i in range(1000):
    #从np.arange(len(x_vals_train))生成大小为20的均匀随机样本,如：[ 66  42  96 115  45 127  31  70 148  57  60 127  56  96   7  63  75 127 110 144]
    rand_index = np.random.choice(len(x_vals_train), size=batch_size)
    #print('rand_index ' + str(rand_index))
    #rand_x为20x2的数组，类似酱紫 [[4.5 1.5] 。。。。[1.3 0.2]]
    rand_x = x_vals_train[rand_index]
    #print(' rand_x ' + str(rand_x))
    #print(' rand_x shape: ' + str(rand_x.shape))

    rand_x1 = np.array([[x[0]] for x in rand_x])
    rand_x2 = np.array([[x[1]] for x in rand_x])
    #print(' rand_x1 ' + str(rand_x1))
    #print(' rand_x2 ' + str(rand_x2))

    #rand_y如果直接使用y_vals_train[rand_index]，则得到的是[0 0 0 0 0 0 1 0 0 0 0 1 1 1 0 1 1 0 0 0]，是一维的, shape是（20,）也就是一维数组，数组有20个元素
    #但是我们想要多维数组，也就是shape为(20, 1)，因为placeholder也是这样的维度
    #[[y] for y in y_vals_train[rand_index]] 得到的是[[0], [0], [0], [0], [0], [0], [1], [0], [1], [0], [0], [0], [0], [1], [1], [1], [1], [1], [0], [0]]
    #然后，转化为数组，维度是(20, 1)
    rand_y = np.array([[y] for y in y_vals_train[rand_index]])
    #print('rand_y shape ' + str(rand_y.shape))
    #print('rand_y  ' + str(rand_y))

    sess.run(train_step, feed_dict={x1_data: rand_x1, x2_data: rand_x2, y_target: rand_y})
    if(i+1)%200==0:
        print('step: ' + str(i) + ' A = ' + str(sess.run(A)) + ' b = ' + str(sess.run(b)))


#-----------------4. 评估模型--------------------------#
#用已有的A,b 变量，计算my_output
y_prediction = tf.squeeze(tf.round(tf.nn.sigmoid(my_output)))
correct_prediction = tf.equal(y_prediction, y_target)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

#测试集数据提取，要转为二维的矩阵
x_vals_test_x1 = np.array([[x[0]] for x in x_vals_test])
x_vals_test_x2 = np.array([[x[1]] for x in x_vals_test])
y_vals_test_predict = np.array([[y] for y in y_vals_test])

#训练集数据提取，要转为二维的矩阵
x_vals_train_x1 = np.array([[x[0]] for x in x_vals_train])
x_vals_train_x2 = np.array([[x[1]] for x in x_vals_train])
y_vals_train_predict = np.array([[y] for y in y_vals_train])

print('x_vals_test ')
print(x_vals_test)
print('x_vals_test_x1 ')
print(x_vals_test_x1)
print('x_vals_test_x2 ')
print(x_vals_test_x2)

print('y_vals_test: ')
print(y_vals_test)
print('y_vals_test_predict: ')
print(y_vals_test_predict)


acc_value_test = sess.run(accuracy, feed_dict={x1_data: x_vals_test_x1, x2_data: x_vals_test_x2, y_target: y_vals_test_predict})
acc_value_train = sess.run(accuracy, feed_dict={x1_data: x_vals_train_x1, x2_data: x_vals_train_x2, y_target: y_vals_train_predict})

print('Accuracy on test set ' + str(acc_value_test))
print('Accuracy on train set ' + str(acc_value_train))