TensorFlow 1.x 基本知识总结

0. name_scope和variable_scope

• name_scope will add scope as a prefix to all operations
• variable_scope will add scope as a prefix to all variables and operations

初始化变量：

• tf.Variable() constructer prefixes variable name with current name_scope and variable_scope
• tf.get_variable() constructor ignores name_scope and only prefixes name with the current variable_scope

使用tf.variable_scope来定义共享变量。

The easiest way to do reuse variables is to use the reuse_variables() as shown below

with tf.variable_scope("scope"):
    var1 = tf.get_variable("variable1",[1])
    tf.get_variable_scope().reuse_variables()
    var2=tf.get_variable("variable1",[1])
assert var1 == var2

https://programming.vip/docs/5d4e8ed7f3666.html

1. placeholder(占位符)

tf.placeholder(
    dtype, #常用的是tf.float32
    shape=None,默认是一维的，如果设成[None, 4]，即为二维的，行数未知
    name=None
)

2. fetch和feed

2.1 fetch可以在一个会话中同时执行多个op

input1 = tf.constant(3.0)
input2 = tf.constant(2.0)
input3 = tf.constant(5.0)

add = tf.add(input2,input3)
mul = tf.multiply(input1,add)

with tf.Session() as sess:
    result = sess.run([mul, add])
    print((result))

2.2 feed可以每次给图传入不同的数据

# 创建占位符
input1 = tf.placeholder(tf.float32)
input2 = tf.placeholder(tf.float32)
output = tf.multiply(input1, input2)

with tf.Session() as sess:
    # feed数据以字典形式传入
    print(sess.run(output, feed_dict={input1:[2.0],input2:[6.0]}))

3. 数学运算相关的API

3.1 tf.nn.embedding_lookup

  以NLP的词嵌入为例，tf.nn.embedding_lookup(word_embedding, input_word)，其中word_embedding的维度是vocab_size（词汇的总个数） * embedding_size（词嵌入的维度），而input_word从一维和二维分别进行说明。

  假设input_word为一维，它的维度为word_num，则结果的维度为word_num * embedding_size，也就是取出了word_num个数个词汇。

  假设input_word为二维，维度是sentence_num(句子的个数) * word_num(词的个数)。则结果的维度为sentence_num * word_num * embedding_size。也就是有sentence_num个句子，其中每个句子为word_num个词，每个词为embedding_size。

import numpy as np
import tensorflow as tf

sess = tf.InteractiveSession()

embedding = tf.Variable(np.identity(6, dtype=np.int32))
input_ids = tf.placeholder(dtype=tf.int32, shape=[None,3])
input_embedding = tf.nn.embedding_lookup(embedding, input_ids)

sess.run(tf.global_variables_initializer())
print (sess.run(embedding))
print ('sepeate paragraph')
print (sess.run(input_embedding, feed_dict={input_ids: [[0, 0, 0], [0, 1, 0]]}))

4. 卷积

4.1 一维卷积

import tensorflow as tf

inputs = tf.placeholder(dtype = 'float', shape=[None, 5, 2])
data = np.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 10]])
data = np.expand_dims(data, axis=0)
print(data.shape)

out = tf.layers.conv1d(inputs, 5, 3) #5个filters,每个filter的windows是3
#print(out)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  print(out)
  print(sess.run(out, feed_dict={inputs:data}))

结果为

(1, 5, 2)
Tensor("conv1d_32/BiasAdd:0", shape=(?, 3, 5), dtype=float32)
[[[ 6.1485815  -1.2904111   8.131317    1.8840454   0.31307578]
  [ 8.398123   -0.84522295 11.939692    3.5162792   0.36305082]
  [10.647663   -0.4000342  15.748065    5.148513    0.41302562]]]