007-Tensor.ipynb · dongzhu1688/tensorflow2_tutorials_chinese

Create your Gitee Account

Explore and code with more than 12 million developers，Free private repositories ！：）

This repository doesn't specify license. Please pay attention to the specific project description and its upstream code dependency when using it.

Clone or Download

007-Tensor.ipynb 9.90 KB

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# TensorFlow2教程-张量\n",
    "\n",
    "由命名可知，TensorFlow是一个定义和运算涉及张量的计算的框架。TensorFlow将张量表示为基本数据类型的n维数组。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "tf.Tensor具有以下属性：\n",
    "- 数据类型（float32, int32, string等）\n",
    "- 形状\n",
    "\n",
    "一个张量中的每个元素都是一样的数据类型，且数据类型已知。形状可能已知也可能部分已知。某些情况下，只有执行图时才知道张量的形状。\n",
    "\n",
    "一些特殊的张量：\n",
    "- tf.Variable\n",
    "- tf.constant\n",
    "- tf.placeholder\n",
    "- tf.SparseTensor\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1 Rank\n",
    "rank是tf.Tensor对象的尺寸的数量。类似于order,degree或ｎ维。TensorFlow中的rank和数据中矩阵的秩不同。如下表所示，TensorFlow每个rank对应一个不同的数学实体。\n",
    "\n",
    "| 秩\t| 数学实体 |\n",
    "| ---- | ---- |\n",
    "| 0\t| 标量（仅幅度） |\n",
    "| 1\t| 向量（大小和方向） |\n",
    "| 2\t| 矩阵（数字表） |\n",
    "| 3\t| 3维张量（数字的立方） |\n",
    "| ñ\t| n维张量 |"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.1 rank 0\n",
    "rank0张量创建"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "mammal = tf.Variable('Hongmeng', tf.string)\n",
    "ignition = tf.Variable(360, tf.int16)\n",
    "floating = tf.Variable(3.141592653, tf.float64)\n",
    "its_complicated = tf.Variable(12.3 - 4.56j, tf.complex64)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "注：在TensorFlow中，字符串被视为单个对象，而不是字符序列。可能有标量字符串，字符串向量等。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.2 rank 1\n",
    "要创建rank 1的tf.Tensor对象，一般传递列表作为初始值"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "mystr = tf.Variable([\"Hello\"], tf.string)\n",
    "cool_numbers  = tf.Variable([3.14159, 2.71828], tf.float32)\n",
    "first_primes = tf.Variable([2, 3, 5, 7, 11], tf.int32)\n",
    "its_very_complicated = tf.Variable([12.3 - 4.85j, 7.5 - 6.23j], tf.complex64)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.3 rank 2\n",
    "rank 2包含行跟列\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "mymat = tf.Variable([[7],[11]], tf.int16)\n",
    "myxor = tf.Variable([[False, True],[True, False]], tf.bool)\n",
    "linear_squares = tf.Variable([[4], [9], [16], [25]], tf.int32)\n",
    "squarish_squares = tf.Variable([ [4, 9], [16, 25] ], tf.int32)\n",
    "rank_of_squares = tf.rank(squarish_squares)\n",
    "mymatC = tf.Variable([[7],[11]], tf.int32)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.4 高阶张量\n",
    "高阶张量由n维数组组成。例如，在图像处理期间，使用了许多等级4的张量，其尺寸对应于批中示例，图像高度，图像宽度和颜色通道。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "my_image = tf.zeros([10, 299, 299, 3]) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.5 获取tf.Tensor对象的rank\n",
    "要确定tf.Tensor对象的rank，可以调用tf.rank方法。例如，以下方法以编程方式确定"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tf.Tensor(4, shape=(), dtype=int32)\n"
     ]
    }
   ],
   "source": [
    "r = tf.rank(my_image)\n",
    "print(r)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.6 tf.Tensor切片\n",
    "\n",
    "由于 tf.Tensor是n维数组，因此要访问其中的某个元素，需要指定n个索引。\n",
    "\n",
    "对于等级0的张量（标量），不需要索引，因为它已经是一个数字。\n",
    "\n",
    "对于秩为1的张量（向量），传递单个索引可访问数字：\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "my_scalar = first_primes[2]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "对于2级或更高的张量，情况更加有趣。对于 tf.Tensor的rank为2的，传递两个数字将按预期返回标量："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "my_scalar = mymat[1, 0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "但是，传递单个数字将返回矩阵的子向量"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "my_row_vector = mymat[1]\n",
    "my_column_vector = mymat[:, 0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2 形状\n",
    "张量的形状是每个维度中元素的数量。TensorFlow在图形构建过程中自动推断形状。这些推断的形状可能具有已知或未知的等级。如果rank是已知的，则每个维度的大小可能是已知的或未知的。\n",
    "\n",
    "TensorFlow文档使用三种符号约定来描述张量维数：rank，形状和维数。下表显示了它们之间的相互关系：\n",
    "\n",
    "|rank\t|形状\t|维数\t|例子|\n",
    "|---- |---- |---- |---- |\n",
    "|0\t|[]\t|0维\t|0维张量。标量。|\n",
    "|1\t|[D0]\t|一维\t|形状为[5]的一维张量。|\n",
    "|2\t|[D0，D1]\t|2维\t|形状为[3，4]的二维张量。|\n",
    "|3\t|[D0，D1，D2]\t|3维\t|形状为[1、4、3]的3-D张量。|\n",
    "|ñ\t|[D0，D1，... Dn-1]\t|D维\t|形状为[D0，D1，... Dn-1]的张量。|"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2.1 获取tf.Tensor对象的形状\n",
    "有两种访问形状tf.Tensor的方法。在构建图形时，询问关于张量形状的已知知识通常很有用。\n",
    "- 1、可以通过读取对象的shape属性来获取tf.Tensor。此方法返回一个TensorShape对象，这是表示部分指定的形状的便捷方式（因为在构建图形时，并非所有形状都将是完全已知的）。\n",
    "\n",
    "- 2、也有可能在运行时获得一个代表某个张量的形状的tf.Tensor。这是通过调用tf.shape 操作来完成的。这样，可以通过构建依赖于输入的动态形状的其他张量来完成网络构建。\n",
    "\n",
    "例如，下面是如何制作一个零向量，其大小与给定矩阵中的列数相同："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "zeros = tf.zeros(mymat.shape[1])\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2.2 改变tf.Tensor的形状\n",
    "可以通过tf.reshape来改变张量的形状，只需确保所有形状的尺寸的乘积相同"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "rank_three_tensor = tf.ones([3, 4, 5])\n",
    "matrix = tf.reshape(rank_three_tensor, [6, 10]) \n",
    "matrixB = tf.reshape(matrix, [3, -1])  .\n",
    "matrixAlt = tf.reshape(matrixB, [4, 3, -1])   \n",
    "                                            "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3 数据类型\n",
    "除维数外，张量还具有数据类型。\n",
    "\n",
    "具有一个以上的数据类型是不可能的。但是，可以将任意数据结构序列化为strings并将其存储在tf.Tensors中。\n",
    "\n",
    "可以使用tf.cast命令将tf.Tensors从一种数据类型转换为另一种数据类型："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "float_tensor = tf.cast(tf.constant([1, 2, 3]), dtype=tf.float32)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "要检查tf.Tensor的数据类型，请使用Tensor.dtype属性。\n",
    "\n",
    "tf.Tensor从python对象创建时，可以选择指定数据类型。如果不设置，TensorFlow会选择一种可以代表数据的数据类型。TensorFlow将Python整数转换为tf.int32，并将python浮点数转换为tf.float32。而在将numpy转换为数组时使用的相同规则。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4 打印张量\n",
    "\n",
    "出于调试目的，可能需要打印tf.Tensor的值。TensorFlow2中可以直接用print打印\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tf.Tensor(7, shape=(), dtype=int32)\n"
     ]
    }
   ],
   "source": [
    "t = tf.add(3, 4)\n",
    "print(t)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

Copy Edit Raw Blame History

authored 2020-02-07 20:16 . 20200207

马建仓 AI 助手

尝试更多

代码解读

代码找茬

代码优化

https://gitee.com/dongzhulab_admin/tensorflow2_tutorials_chinese.git

git@gitee.com:dongzhulab_admin/tensorflow2_tutorials_chinese.git

dongzhulab_admin

tensorflow2_tutorials_chinese

master

dongzhu1688 / tensorflow2_tutorials_chinese

About

Releases

Contributors

Activities