python3 + tensorflow1.8.0

基于Github上的例子，不定时更新



两种打开session方式等价
[python] view plain copy
import tensorflow as tf  
y=tf.constant(23)  
with tf.Session() as sess: #会自动关闭session  
    print(sess.run(y))  
  
sess = tf.Session()  
print(sess.run(y))  
sess.close()  
这两种方式需要在启动session之前构建完计算图

更灵活的方式是

[python] view plain copy
sess = tf.InteractiveSession()  

常数和变量
[python] view plain copy
#常数  
y = tf.constant(23,name='y')  
#变量  
y2 = tf.Variable(23,name='y2')  
y3 = tf.placeholder(tf.int32) #占位符，类似形参  
  
init = tf.global_variables_initializer() #初始化节点  
sess = tf.Session()  
sess.run(init)  
print(sess.run(y),sess.run(y2))  
  
z = tf.identity(y3) #占位符无法引用自身，放入计算图至少需要一个操作  
value = 23  
print(sess.run(z,feed_dict={y3:value}))  
设置参数
[python] view plain copy
#第一个是参数名称，第二个参数是默认值，第三个是参数描述  
tf.app.flags.DEFINE_string('str_name', 'value',"descrip1")  
FLAGS = tf.app.flags.FLAGS  
  
加法，乘法
[python] view plain copy
x = tf.placeholder(tf.int32)  
y = tf.placeholder(tf.int32)  
add = tf.add(x, y)  
mul = tf.multiply(x, y)  
  
with tf.Session() as sess:  
    print(sess.run(add, feed_dict={x: 2, y: 3}))  
    print(sess.run(mul, feed_dict={x: 2, y: 3}))  
  

矩阵运算
[python] view plain copy
matrix1 = tf.constant([[3., 3.]])  
matrix2 = tf.constant([[2.],[2.]])  
product = tf.matmul(matrix1, matrix2)  
add = tf.add(matrix1, matrix2)  
with tf.Session() as sess:  
    result1 = sess.run(product)  
    result2 = sess.run(add)  
    print(result1,result2)  
  
线性函数
[python] view plain copy
def linear_function():  
    np.random.seed(1)  
    X = tf.constant(np.random.randn(3,1),name='X')  
    W = tf.constant(np.random.randn(4,3),name='W')  
    b = tf.constant(np.random.randn(4,1),name='b')  
    Y = tf.add(tf.matmul(W,X),b)  
    a = tf.constant(2)  
    sess = tf.Session()  
    result = sess.run(Y)  
    sess.close()  
    return result  
print(str(linear_function()))  
  

sigmoid函数
[python] view plain copy
def sigmoid(z):  
    x = tf.placeholder(tf.float32,name = "x")  
    sigmoid = tf.sigmoid(x)  
    with tf.Session() as sess:  
        result = sess.run(sigmoid,feed_dict = {x:z})  
    return result  
print(sigmoid(2))  
  

转化为一个one hot向量
[python] view plain copy
def one_hot_matrix(labels,C):  
    depth = tf.constant(value=C,name='C')  
    one_hot_matrix = tf.one_hot(labels,depth,axis=0)  
    sess = tf.Session()  
  
    one_hot = sess.run(one_hot_matrix)  
    sess.close()  
    return one_hot  
  
labels = np.array([1,2,3,0,2,1])  
one_hot = one_hot_matrix(labels, C = 4)  
#4x6矩阵，四行分别对应0，1，2，3；元素为1表示出现的位置  
print (str(one_hot))   
  
保存模型
[python] view plain copy
# 每训练50次保存一次模型  
if i% 50 == 0:  
    saver.save(sess, self.ckpt_path + self.model_name + '.ckpt', global_step=i)  
会自动存为三个文件：

model_name.ckpt-iteration_step.data-00000-of-00001, model_name.ckpt-iteration_step.index, model_name.ckpt-iteration_step.meta

以及一个chekpoint文件保存最新模型的名字，内容：

model_checkpoint_path: "model_name.ckpt-iteration_step"

all_model_checkpoint_paths: "model_name.ckpt-0"

  

读取之前保存的模型
[python] view plain copy
def restore_last_session(ckpt_path):  
    saver = tf.train.Saver()  
    sess = tf.Session()  
    ckpt = tf.train.get_checkpoint_state(ckpt_path)  
    if ckpt and ckpt.model_checkpoint_path:  
        saver.restore(sess, ckpt.model_checkpoint_path)  
    return sess  
[python] view plain copy
sess=restore_last_session('yourpath')  
ckpt.model_checkpoint_path会自动在checkpoint文件读取最新模型的名字
  

线性拟合
[python] view plain copy
import tensorflow as tf  
import numpy  
import matplotlib.pyplot as plt  
rng = numpy.random  
  
# Parameters  
learning_rate = 0.01  
training_epochs = 2000  
display_step = 50  
  
# Training Data  
train_X = numpy.asarray([3.3,4.4,5.5,6.71,6.93,4.168,9.779,6.182,7.59,2.167,7.042,10.791,5.313,7.997,5.654,9.27,3.1])  
train_Y = numpy.asarray([1.7,2.76,2.09,3.19,1.694,1.573,3.366,2.596,2.53,1.221,2.827,3.465,1.65,2.904,2.42,2.94,1.3])  
n_samples = train_X.shape[0]  
  
# tf Graph Input  
X = tf.placeholder("float")  
Y = tf.placeholder("float")  
  
# Create Model  
  
# Set model weights  
W = tf.Variable(rng.randn(), name="weight")  
b = tf.Variable(rng.randn(), name="bias")  
  
# Construct a linear model  
activation = tf.add(tf.multiply(X, W), b)  
  
# Minimize the squared errors  
cost = tf.reduce_sum(tf.pow(activation-Y, 2))/(2*n_samples) #L2 loss  
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost) #Gradient descent  
  
# Initializing the variables  
init = tf.initialize_all_variables()  
  
# Launch the graph  
with tf.Session() as sess:  
    sess.run(init)  
  
    # Fit all training data  
    for epoch in range(training_epochs):  
        for (x, y) in zip(train_X, train_Y):  
            sess.run(optimizer, feed_dict={X: x, Y: y})  
  
        #Display logs per epoch step  
        if epoch % display_step == 0:  
            print("Epoch:", '%04d' % (epoch+1), "cost=", \  
                "{:.9f}".format(sess.run(cost, feed_dict={X: train_X, Y:train_Y})), \  
                "W=", sess.run(W), "b=", sess.run(b) )  
  
    print("Optimization Finished!")  
    print("cost=", sess.run(cost, feed_dict={X: train_X, Y: train_Y}), \  
          "W=", sess.run(W), "b=", sess.run(b))  
  
    #Graphic display  
    plt.plot(train_X, train_Y, 'ro', label='Original data')  
    plt.plot(train_X, sess.run(W) * train_X + sess.run(b), label='Fitted line')  
    plt.legend()  
    plt.show()