TensorFlow1.0 - C2 Guide - 4 Low

1 Intro

组成 tf.Graph 和 tf.Session
简单的输入constant，placeholder，feeding，Datasets
layers和feature columns和初始化方法
loss和optimizer来训练

• Abstract
  Use tf.Graph an tf.Session
  Use high level components(datasets, layers and feature_columns
  Build your own training loop

• Core walkthrough
  A graph consists of two parts: tf.Operation and tf.Tensor

• TensorBoard
  tf.summary.FileWriter('.')
  https://www.tensorflow.org/guide/summaries_and_tensorboard

• Session

sess = tf.Session()
sess.run({'ab':(a, b), 'total':total})

tf.random_uniform每次同时产生的是一个数据

• Feeding

x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
z = x + y
print(sess.run(z, feed_dict={x: 3, y: 4.5}))
print(sess.run(z, feed_dict={x: [1, 3], y: [2, 4]}))

• Datasets
  tf.data.Dataset.from_tensor_slices(my_data)
  next_item = slices.make_one_shot_iterator().get_next()
  主要是可以有iterator机制
• Layers
  主要有两种写法，一种是tf.layers.Dense(units=1)和函数式写法tf.layers.dense(x, units=1)
  运行时需要init = tf.global_variables_initializer()初始化weight
• Feature columns
  用tf.feature_column.input_layer实现
  分为numeric_column和indicator_column
  有时候feature columns有内部状态，需要像layers一样被初始化，categorical用的是tf.contrib.lookup，用tf.tables_initializer初始化
• Training
  定义一个loss
  定义optimizer
  然后组合训练

optimizer = tf.train.GradientDescentOptimizer(0.01)
train = optimizer.minimize(loss)

完整代码

x = tf.constant([[1], [2], [3], [4]], dtype=tf.float32)
y_true = tf.constant([[0], [-1], [-2], [-3]], dtype=tf.float32)

linear_model = tf.layers.Dense(units=1)

y_pred = linear_model(x)
loss = tf.losses.mean_squared_error(labels=y_true, predictions=y_pred)

optimizer = tf.train.GradientDescentOptimizer(0.01)
train = optimizer.minimize(loss)

init = tf.global_variables_initializer()

sess = tf.Session()
sess.run(init)
for i in range(100):
  _, loss_value = sess.run((train, loss))
  print(loss_value)

print(sess.run(y_pred))

2 Tensors

有两个属性data type和shape
有四种类型
rank和shape

• Rank
  就是维度
  获得维度 r = tf.rank(my_image)
• Shape
  表示每一个维度的个数
  获得shape，zeros = tf.zeros(my_matrix.shape[1])
  更改shape

rank_three_tensor = tf.ones([3, 4, 5])
matrix = tf.reshape(rank_three_tensor, [6, 10])  # Reshape existing content into
                                                 # a 6x10 matrix
matrixB = tf.reshape(matrix, [3, -1])  #  Reshape existing content into a 3x20
                                       # matrix. -1 tells reshape to calculate
                                       # the size of this dimension.
matrixAlt = tf.reshape(matrixB, [4, 3, -1])  # Reshape existing content into a
                                             #4x3x5 tensor

• Data types
  tf.string, tf.int32, tf.float32，后面两个是python的默认值
• Printing Tensors
  Evaluating Tensors
  Print Tensors

3 Variables

创建变量
collection和device placement
使用和初始化变量
共享变量

• Creating a variable

my_variable = tf.get_variable("my_variable", [1, 2, 3])

会有默认初始化方法，也可以设置初始化方法

• variable collections
  为了在一个地方使用所有的变量
  默认有两个collection：

• tf.GraphKeys.GLOBAL_VARIABLES --- variables that can be shared across multiple devices,
• tf.GraphKeys.TRAINABLE_VARIABLES --- variables for which TensorFlow will calculate gradients.
  变量可以设置是否trainable和在哪个collections

• device placement
  可以把变量放在particular devices
• Initializing
  高阶api如Estimator,keras会自动初始化

session.run(tf.global_variables_initializer())
# Now all variables are initialized.
session.run(my_variable.initializer)
print(session.run(tf.report_uninitialized_variables()))

一个变量初始化依赖别的变量

v = tf.get_variable("v", shape=(), initializer=tf.zeros_initializer())
w = tf.get_variable("w", initializer=v.initialized_value() + 1)

• Using
  使用正常像tensor一样
  赋值用v.assign()和v.assign_add(1)
  tf.Variable.read_value去in time读变量
• Sharing
  显示sharing，直接传递
  隐示sharing，使用tf.variable_scope("model", reuse=True) reuse来控制是否重载使用

4 Graphs and Sessions

• why and what is tf.Graph
  why: parallelism distributed compilation portability
  tf.Graph contains graph structure and graph collections
• Building a tf.Graph

5 Save and Restore

6 Control Flow

7 Ragged Tensors