Summary of tensorflow basic knowledge + linear regression practice

1. Introduction to Tensorflow

1.2 Graph and tensorBoard

Achieving program visualization requires two steps:

1) Data serialization-events file

import tensorflow as tf
a = tf.constant(20)
b = tf.constant(22)
c = tf.add(a, b)

# 获取默认图
g = tf.get_default_graph()
print("图", g)

with tf.Session() as sess:
    # 1、写入到events文件中
    file_writer = tf.summary.FileWriter("./temp/summary", graph=sess.graph)
    sum_t = sess.run(c)
    print(sum_t)

2) Start TensorBoard

tensorboard --logdir="./ai/temp/summary/" --host=127.0.0.1  # 在cmd中输入自己的events文件所在目录

> TensorBoard 1.12.2 at http://127.0.0.1:6006 (Press CTRL+C to quit)

Enter in the browser: http://127.0.0.1:6006 to openTensorBoard

2.on

2.1 Modify the naming of tensors

a = tf.constant(20, name="a")
b = tf.constant(20, name="b")

3. Conversation

3.1 __ heat __

import tensorflow as tf
a = tf.constant(20, name="a")
b = tf.constant(22, name="b")
d = tf.constant(22.2)
c = tf.add(a, b)

# 获取默认图
g = tf.get_default_graph()
print("图", g)
print("a:", a)
print("d:", d)

with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)) as sess:
    # 1、写入到events文件中
    file_writer = tf.summary.FileWriter("./temp/summary", graph=sess.graph)
    sum_t = sess.run(c)
    print(sum_t)

>
图 <tensorflow.python.framework.ops.Graph object at 0x000001D2D0B02470>
a: Tensor("a:0", shape=(), dtype=int32)
d: Tensor("Const:0", shape=(), dtype=float32)

Device mapping: no known devices.
Add: (Add): /job:localhost/replica:0/task:0/device:CPU:0
a: (Const): /job:localhost/replica:0/task:0/device:CPU:0
b: (Const): /job:localhost/replica:0/task:0/device:CPU:0
Const: (Const): /job:localhost/replica:0/task:0/device:CPU:0
42

3.2 run() in session

import tensorflow as tf
a = tf.constant(20, name="a")
b = tf.constant(22, name="b")
d = tf.constant(22.2)
c = tf.add(a, b)


with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)) as sess:
    # 1、写入到events文件中
    file_writer = tf.summary.FileWriter("./temp/summary", graph=sess.graph)
    a, b, c = sess.run([a, b, c])
    print(a, b, c)

feed_dict: Ability to control variable assignment at runtime

import tensorflow as tf


# 定义placeholder
plat_a = tf.placeholder(dtype=tf.float32)
plat_b = tf.placeholder(dtype=tf.float32)
plat_add = tf.add(plat_a, plat_b)

with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)) as sess:
    # 1、写入到events文件中
    file_writer = tf.summary.FileWriter("./temp/summary", graph=sess.graph)
    ret = sess.run(plat_add, feed_dict={
    
    plat_a: 5.0, plat_b: 6.0})
    print(ret)

4 tensors

TensorFlow's tensor is an n-dimensional array of type tf.Tensor. Tensor has two important properties:

• type:数据类型
• shape:形状（阶）

4.1 Types of tensors

 ones = tf.ones([3, 4])
 print(ones.eval())

4.2 Transformation of tensor shape

4.3 Shape change

TensorFlow's tensors have two shape transformations, dynamic shape and static shape:

• tf.reshape
• tf.set_shape
  
  Static shape settings:

# 定义placeholder
plat_a = tf.placeholder(dtype=tf.float32)
plat_b = tf.placeholder(dtype=tf.float32)
plat_add = tf.add(plat_a, plat_b)

plat_c = tf.placeholder(dtype=tf.float32, shape=[None, None])
print(plat_c.shape)
plat_c.set_shape([5, 6])
print("c.shape:", plat_c.shape)
plat_c.set_shape([4, 3])  # 不能再次设置静态形状，否则会报错!!!，也不能跨阶数修改！！！！
print("c.shape:", plat_c.shape)

Dynamic shape settings: a new tensor value will be generated

plat_c = tf.placeholder(dtype=tf.float32, shape=[None, None])
print(plat_c.shape)
c_shape = tf.reshape(plat_c, [12, 12])
c_shape = tf.reshape(c_shape, [2, 12, 6])
print("c.shape:", c_shape.get_shape())

4.4 Mathematical operations on tensors

4.5 Variable OP

2. Use TensorFlow to implement linear regression case

2.1 Basic code

import tensorflow as tf


# 1.准备数据
X = tf.random_normal(shape=(100, 1), mean=2, stddev=2)
y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
weights = tf.Variable(initial_value=tf.random_normal((1, 1)))
bias = tf.Variable(initial_value=tf.random_normal((1, 1)))
y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)


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

# 5.训练
with tf.Session() as sess:
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))
    # 训练模型
    for i in range(1000):
        sess.run(optimizer)
        print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))

2.2 Namespace

Add module naming

import tensorflow as tf


# 1.准备数据
with tf.variable_scope("original_data"):
    X = tf.random_normal(shape=(100, 1), mean=2, stddev=2, name="original_data_x")
    y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
with tf.variable_scope("linear_model"):
    weights = tf.Variable(initial_value=tf.random_normal((1, 1)), name="w")
    bias = tf.Variable(initial_value=tf.random_normal((1, 1)), name="b")
    y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
with tf.variable_scope("loss"):
    error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
with tf.variable_scope("optimizer"):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)


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

# 5.训练
with tf.Session() as sess:
    # 5.1运行初始化变量
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))
    
    # 5.2 创建事件文件
    file_writer = tf.summary.FileWriter(logdir="./temp/summary/", graph=sess.graph)

    # 5.3训练模型
    for i in range(1000):
        sess.run(optimizer)
        print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))

2.3 Collect variables

import tensorflow as tf


# 1.准备数据
with tf.variable_scope("original_data"):
    X = tf.random_normal(shape=(100, 1), mean=2, stddev=2, name="original_data_x")
    y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
with tf.variable_scope("linear_model"):
    weights = tf.Variable(initial_value=tf.random_normal((1, 1)), name="w")
    bias = tf.Variable(initial_value=tf.random_normal((1, 1)), name="b")
    y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
with tf.variable_scope("loss"):
    error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
with tf.variable_scope("optimizer"):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)

# 11.手机观察的变量
tf.summary.scalar("losses", error)
tf.summary.histogram("weight", weights)
tf.summary.histogram("biases", bias)

# 22.合并收集的张量
merge = tf.summary.merge_all()

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

# 5.训练
with tf.Session() as sess:
    # 5.1运行初始化变量
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))

    # 5.2 创建事件文件
    file_writer = tf.summary.FileWriter(logdir="./temp/summary/", graph=sess.graph)

    # 5.3训练模型
    for i in range(1000):
        sess.run(optimizer)
        print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))

        # 写入收集的变量
        summary = sess.run(merge)
        file_writer.add_summary(summary, i)

2.4 Model saving and loading

The saved file format is ckpt
saver = tf.train.Saver()
saver.save(sess, "./temp/ckpt/linearregression")
saver.restore(sess, "./temp/ckpt/linearregression")

save code

import tensorflow as tf


# 1.准备数据
with tf.variable_scope("original_data"):
    X = tf.random_normal(shape=(100, 1), mean=2, stddev=2, name="original_data_x")
    y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
with tf.variable_scope("linear_model"):
    weights = tf.Variable(initial_value=tf.random_normal((1, 1)), name="w")
    bias = tf.Variable(initial_value=tf.random_normal((1, 1)), name="b")
    y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
with tf.variable_scope("loss"):
    error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
with tf.variable_scope("optimizer"):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)

# 11.手机观察的变量
tf.summary.scalar("losses", error)
tf.summary.histogram("weight", weights)
tf.summary.histogram("biases", bias)

# 22.合并收集的张量
merge = tf.summary.merge_all()

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

# 创建saver
saver = tf.train.Saver()

# 5.训练
with tf.Session() as sess:
    # 5.1运行初始化变量
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))

    # 5.2 创建事件文件
    file_writer = tf.summary.FileWriter(logdir="./temp/summary/", graph=sess.graph)

    # 5.3训练模型
    for i in range(1000):
        sess.run(optimizer)
        print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))

        # 写入收集的变量
        summary = sess.run(merge)
        file_writer.add_summary(summary, i)

        saver.save(sess, "./temp/ckpt/linearregression")

read

import tensorflow as tf


# 1.准备数据
with tf.variable_scope("original_data"):
    X = tf.random_normal(shape=(100, 1), mean=2, stddev=2, name="original_data_x")
    y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
with tf.variable_scope("linear_model"):
    weights = tf.Variable(initial_value=tf.random_normal((1, 1)), name="w")
    bias = tf.Variable(initial_value=tf.random_normal((1, 1)), name="b")
    y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
with tf.variable_scope("loss"):
    error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
with tf.variable_scope("optimizer"):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)

# 11.手机观察的变量
tf.summary.scalar("losses", error)
tf.summary.histogram("weight", weights)
tf.summary.histogram("biases", bias)

# 22.合并收集的张量
merge = tf.summary.merge_all()

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

# 创建saver
saver = tf.train.Saver()

# 5.训练
with tf.Session() as sess:
    # 5.1运行初始化变量
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))

    # 5.2 创建事件文件
    file_writer = tf.summary.FileWriter(logdir="./temp/summary/", graph=sess.graph)

    saver.restore(sess, "./temp/ckpt/linearregression")
    print(weights.eval(), bias.eval())
    # # 5.3训练模型
    # for i in range(1000):
    #     sess.run(optimizer)
    #     print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))
    #
    #     # 写入收集的变量
    #     summary = sess.run(merge)
    #     file_writer.add_summary(summary, i)
    #
    #     saver.save(sess, "./temp/ckpt/linearregression")

2.4 Command line parameters

import tensorflow as tf


# 1.准备数据
with tf.variable_scope("original_data"):
    X = tf.random_normal(shape=(100, 1), mean=2, stddev=2, name="original_data_x")
    y_true = tf.matmul(X, [[0.8]]) + [[0.7]]

# 2.构建模型
with tf.variable_scope("linear_model"):
    weights = tf.Variable(initial_value=tf.random_normal((1, 1)), name="w")
    bias = tf.Variable(initial_value=tf.random_normal((1, 1)), name="b")
    y_predict = tf.matmul(X, weights) + bias

# 3. 确定损失函数
with tf.variable_scope("loss"):
    error = tf.reduce_mean(tf.square(y_predict-y_true))

# 4.梯度下降优化损失
with tf.variable_scope("optimizer"):
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(error)

# 11.手机观察的变量
tf.summary.scalar("losses", error)
tf.summary.histogram("weight", weights)
tf.summary.histogram("biases", bias)

# 22.合并收集的张量
merge = tf.summary.merge_all()

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

# 创建saver
saver = tf.train.Saver()

# ####################  定义命令行参数
tf.app.flags.DEFINE_integer("max_step", 1000, "train_ step number")
FLAGS = tf.app.flags.FLAGS

# 5.训练
with tf.Session() as sess:
    # 5.1运行初始化变量
    sess.run(init)
    print("随机初始化的权重%f, 偏置为%f" % (weights.eval(), bias.eval()))

    # 5.2 创建事件文件
    file_writer = tf.summary.FileWriter(logdir="./temp/summary/", graph=sess.graph)

    # saver.restore(sess, "./temp/ckpt/linearregression")
    print(weights.eval(), bias.eval())
    # 5.3训练模型
    for i in range(FLAGS.max_step):
        sess.run(optimizer)
        print("第%d步的误差为%f, 权重为%f, 偏置为%f" % (i, error.eval(), weights.eval(), bias.eval()))

        # 写入收集的变量
        summary = sess.run(merge)
        file_writer.add_summary(summary, i)

        saver.save(sess, "./temp/ckpt/linearregression")