要确保已经明白神经网络和卷积神经网络的原理.如果不明白,先学习参考资料1.tensorflow中有很多api,可以分成2大类.1类是比较低层的api(tf.train),叫TensorFlow Core.还有1种相对高层的api(tf.contrib.learn),是建立在TensorFlow Core基础上的,这种api码农用着很方便.

环境

python 3.5.3
tensorflow 1.0.0

Tensors

TensorFlow中的基本数据是tensor. tensor可以直观地理解为把numpy中的数组又包了一层.tensor的runk表示tensor是几维的.比如

[1. ,2., 3.] # runnk为1的tensor,它的shape是[3]
[[1., 2., 3.], [4., 5., 6.]] # runnk为2的tensor,它的shape是[2, 3]
[[[1., 2., 3.]], [[7., 8., 9.]]] # runnk为3的tensor,它的shape是[2, 1, 3]

helloworld级使用

使用tensorflow编程有2个步骤.第1是建立computational graph,第2是运行computational graph.computational graph中的每个结点都有0个或多个tensor作为输入.有一种结点本身是个常量,这种结点没有输入,有固定的输出(即它本身).下面是2个结点:

node1 = tf.constant(3.0, tf.float32)
node2 = tf.constant(4.0) # 默认类型就是tf.float32
print(node1, node2)

运行这个代码结果是

Tensor("Const:0", shape=(), dtype=float32) Tensor("Const_1:0", shape=(), dtype=float32)

注意输出中没有具体的值3.0,4.0.这可以理解成建立computational graph.在通过Session运行computational graph的时候才会把值填到结点中.比如

node1 = tf.constant(3.0, tf.float32)
node2 = tf.constant(4.0) # 默认类型就是tf.float32
sess = tf.Session()
print(sess.run([node1, node2]))

稍微复杂一点的例子.

node1 = tf.constant(3.0, tf.float32)
node2 = tf.constant(4.0) # 默认类型就是tf.float32
node3 = tf.add(node1, node2)
print("node3: ", node3)
sess = tf.Session()
print("sess.run(node3): ",sess.run(node3))

placeholder

placeholder有什么作用?placeholder可以用来先定义一种操作,执行的时候再具体赋值.比如python函数的定义

def add(a, b)
    return a + b

a,b都没有具体的值,调用的时候才赋值.不严谨但直观地可以把a,b理解为placeholder.下面看tensorflow的placeholder.

a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
adder_node = a + b
print(sess.run(adder_node, {a: 3, b: 4.5}))  # 输出为7.5
print(sess.run(adder_node, {a: [1,3], b: [2, 4]}))  # 输出为[ 3.  7.]

再看一个例子

import tensorflow as tf
a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
adder_node = a + b
add_and_triple = adder_node * 3
sess = tf.Session()
print(sess.run(add_and_triple, {a: 3, b: 4.5}))  # 输出为22.5

Variable

可以简单地认为在训练的各个参数即为Variable.看下面的例子.

import tensorflow as tf
W = tf.Variable([.3], tf.float32)
b = tf.Variable([-.3], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
print(sess.run(linear_model, {x: [1, 2, 3, 4]}))

输出为W * x + b的值.下面看怎么使用损失函数.

import tensorflow as tf

b = tf.Variable([-.3], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
y = tf.placeholder(tf.float32)
squared_deltas = tf.square(linear_model - y)
loss = tf.reduce_sum(squared_deltas)
print(sess.run(loss, {x: [1, 2, 3, 4], y: [0, -1, -2, -3]}))  # 损失函数是23.66

修改w,b的值再看下损失函数.

import tensorflow as tf
W = tf.Variable([.3], tf.float32)
b = tf.Variable([-.3], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)

y = tf.placeholder(tf.float32)
squared_deltas = tf.square(linear_model - y)
loss = tf.reduce_sum(squared_deltas)

fixW = tf.assign(W, [-1.])
fixb = tf.assign(b, [1.])
sess.run([fixW, fixb])
print(sess.run(loss, {x: [1, 2, 3, 4], y:[0, -1, -2, -3]}))  # 损失函数是0

训练方法

现在要解决如下问题:
已知向量x=(1, 2, 3, 4),向量y=(0, -1, -2, -3),w,b是标量.求w,b使y=wx+b

import tensorflow as tf

W = tf.Variable([.3], tf.float32)
b = tf.Variable([-.3], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
y = tf.placeholder(tf.float32)
loss = tf.reduce_sum(tf.square(linear_model - y))
optimizer = tf.train.GradientDescentOptimizer(0.01)
train = optimizer.minimize(loss)
x_train = [1,2,3,4]
y_train = [0,-1,-2,-3]
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
for i in range(1000):
    sess.run(train, {x: x_train, y: y_train})
    # print(sess.run(loss, {x: x_train, y: y_train}))输出loss

curr_W, curr_b, curr_loss = sess.run([W, b, loss], {x: x_train, y: y_train})
print("W: %s b: %s loss: %s"%(curr_W, curr_b, curr_loss))

上面代码用梯度下降方法求出w,b.现在来验证下wx+b和y相差多少.

import tensorflow as tf

W = tf.constant([-0.9999969], tf.float32)
b = tf.constant([0.99999082], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
print(sess.run(linear_model, {x: [1, 2, 3, 4]}))

输出为

[ -6.07967377e-06  -1.00000298e+00  -1.99999988e+00  -2.99999666e+00]

已经相当接近(0, -1, -2, -3).
上面是用TensorFlow Core中的方法训练.也可以用较高层的api(tf.contrib.learn). 因为这种方法过于抽象,反而会分散初学都注意力.以后再补上.

问题

• 用较高层的api(tf.contrib.learn)训练

参考资料

1 http://cs231n.github.io/
2 tensorflow官方教程